IJACSA Volume 17 Issue 2

Copyright Statement: This is an open access publication licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, even commercially as long as the original work is properly cited.

View Full Issue

Paper 1: A Comparison of Metaheuristic Methods for the Vehicle Routing Problem

Abstract: The Capacitated Vehicle Routing Problem (CVRP) is a fundamental NP-hard combinatorial optimization problem with important applications in logistics and distribution systems. Although numerous advanced approaches have been proposed in recent years, systematic benchmarking of classical metaheuristic algorithms under a unified experimental framework remains limited. This study evaluates the performance and trade-offs of four well-known metaheuristics: Hill Climbing (HC), Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO), and Genetic Algorithms (GA). All methods are implemented within the same computational environment and assessed on benchmark CVRP instances, using the CPLEX exact solver as a reference for global optimality. The results indicate that ACO achieves the smallest optimality gaps and often approaches optimal solutions, at the cost of higher computational effort. PSO strikes a favorable balance between solution quality and runtime across the tested instances, whereas HC delivers very fast solutions but degrades as problem complexity increases. GA exhibits higher variability and less competitive performance under the selected parameter settings. Overall, this comparative analysis highlights the strengths and limitations of classical metaheuristics and establishes a reproducible baseline for future research, including hybrid and learning-assisted approaches for scalable vehicle routing optimization.

Author 1: Manal El Jaouhari

Author 2: Ghita Bencheikh

Author 3: Ghizlane Bencheikh

Keywords: Metaheuristics; Ant Colony Optimization; Hill Climbing; Genetic Algorithm; Particle Swarm Optimization; exact algorithm; Capacitated Vehicle Routing Problem

Paper 2: A Confidence-Aware Multi-Layer Framework for Drone Forensic Inference Using Heterogeneous Digital Evidence Sources and Flight Logs Analysis

Abstract: The increasing use of unmanned aerial vehicles (UAVs) in criminal and adversarial contexts has created new challenges for digital forensic investigations. Current UAV forensic research primarily emphasizes artefact extraction and platform-dependent analysis, while insufficient attention has been given to uncertainty modelling and confidence quantification in forensic inference. This study addresses this methodological gap by proposing a confidence-aware multi-layer UAV forensic framework designed to support legally defensible forensic conclusions. The framework integrates chip-off memory acquisition, logical flight log analysis, companion mobile device artefact examination, and wireless trace correlation within a unified analytical architecture. Physics-based flight trajectory reconstruction and cross-device temporal alignment algorithms enhance reproducibility and platform independence. To reflect varying levels of evidentiary reliability, a structured evidence-weighting approach is introduced alongside a novel Forensic Confidence Index (FCI) that quantifies evidentiary support without implying absolute certainty. Validation using a Yuneec Typhoon Q500 4K dataset demonstrates feasible trajectory reconstruction, temporal correlation, and confidence-constrained attribution under realistic investigative conditions. By explicitly incorporating uncertainty modeling and confidence articulation into UAV forensic workflows, the proposed framework improves scientific rigor, transparency, and legal defensibility while providing a scalable foundation for future cyber-physical forensic investigations.

Author 1: Nidhiba Parmar

Author 2: Naveen Kumar Chaudhary

Keywords: UAV forensics; drone investigations; forensic inference; confidence quantification; flight log analysis; digital forensic framework

Paper 3: Real-Time Data-Driven Decision Support in Retail: A Hybrid GraphSAGE+XGBoost Model for Predicting Reorder Behavior and Unraveling Consumer Communities

Abstract: The rising demand for real-time, data-driven decision support in retail platforms has underscored the need for intelligent systems capable of modeling both behavioral sequences and product relationships. This study introduces a hybrid architecture for real-time decision support in retailing by coupling graph-based learning with conventional machine learning methods. Based on Instacart 2017 data, it constructs a heterogeneous user-product graph and utilizes GraphSAGE to obtain relational embeddings. This combination of embeddings and domain-specific features is then fed into an XGBoost classifier to predict reorder behavior. Empirical findings show that the proposed GraphSAGE+XGBoost model outperforms conventional baselines, including the sole XGBoost, Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) models. In particular, the hybrid model outperformed all baselines across all metrics, achieving a precision of 0.82, a recall of 0.78, an F1-score of 0.76, and a mean Average Precision (mAP) of 0.75. Furthermore, within the co-purchase network, product-level community identification identified significant clusters (such as breakfast staples, health-conscious products, and impulsive snacking) that provided insights into customer demographics and marketing potential. The experimental analysis comparing the proposed GraphSAGE+XGBoost with baseline models, including LSTM, XGBoost, and MLP, demonstrates that the proposed hybrid model outperforms in terms of modeling accuracy, Precision, and generalizability. The system is optimized for real-time inference and can operate in a dynamic commercial landscape, unraveling complex co-purchase behavior and hidden consumer communities.

Author 1: Balayet Hossain

Author 2: Md Deluar Hossen

Author 3: Md Nuruzzaman Pranto

Author 4: Belal Hossain

Author 5: Sabrina Shamim Moushi

Author 6: Nusrat Ameri

Author 7: Khandakar Rabbi Ahmed

Keywords: Real-time business intelligence; graph analytics; machine learning; GraphSAGE; XGBoost; retail analytics; recommendation systems; decision support; instacart dataset; hybrid model

Paper 4: Evaluating the Efficiency of LLM-Generated Software in Resisting Malicious Attacks

Abstract: This study introduces a structured framework for evaluating the security of Java applications generated by large language models (LLMs) and presents the results from its implementation across three models: DeepSeek, GPT-4, and Llama 4. The framework integrates Open Web Application Security Project (OWASP)-supported tools, such as SpotBugs with FindSecBugs, OWASP Dependency Check, and OWASP Zed Attack Proxy (ZAP), alongside the NIST Risk Management Framework. These tools and standards were selected for being publicly available, allowing this process to be replicated and extended without proprietary licensing, and for their alignment with widely adopted industry benchmarks. The testing methodology for generated Java applications includes static code analysis, third-party dependency checking, and dynamic attack simulation. Each of the specified tools for this study corresponds to identifying a specific category of critical vulnerabilities. Identified vulnerabilities are then evaluated against NIST risk analysis standards to characterize their threat sources, likelihoods, and impacts, as well as their implications for the overall security risk profile of each application. The effect of prompt design is also explored by comparing a neutral prompt against a security-emphasized prompt incorporating OWASP best practices. Results varied considerably across models: GPT-4 showed noticeable improvements across critical and high-severity vulnerabilities, with 33.3% and 53.8% reductions, respectively. However, Llama 4 and DeepSeek saw an increase in vulnerabilities from the neutral to the secure prompt. Llama 4 had a general increase of 10- 15% across critical, high, and medium-severity vulnerabilities, while DeepSeek saw no change in high-severity vulnerabilities and a 40% increase in low-severity vulnerabilities. The framework presented provides a structured process for evaluating LLM-generated code against established software development and security standards, while identifying present limitations and possible directions for future work.

Author 1: Dominic Niceforo

Author 2: Haydar Cukurtepe

Keywords: Artificial intelligence; cyber security; large language models (LLMs); software security evaluation

Paper 5: Beyond the Interface: AI Integration Through Input Stream Mediation and Intelligent Output Simulation

Abstract: As Natural Language Processing (NLP) technologies continue to advance, their integration within everyday software tools remains fragmented and often constrained by environment-specific plugins or proprietary AI interfaces. This study introduces a lightweight, platform-independent framework that transforms any text-input surface, ranging from simple editors to browsers and full-featured integrated development environments (IDEs), into an intelligent, AI-assisted interface. Leveraging keystroke dynamics and key-chord recognition, the system operates unobtrusively in the background to interpret user intent and enable real-time interaction within active applications. It provides context-aware suggestions, completions, and insights directly within the active window, effectively turning ordinary typing environments into responsive, intelligent companions. Beyond unifying AI support across diverse applications, the framework enables users to seamlessly switch among multiple open-source AI and NLP models via simple key combination triggers, supported through flexible API integration, thereby providing dynamic access to different linguistic capabilities on demand. The architecture also incorporates a protective intelligence layer that detects suspicious behavior patterns and safeguards sensitive information, offering an additional shield against unauthorized data exposure. By employing a universal mediation layer for input and output, supported by controlled keystroke simulation, the approach eliminates the need for custom extensions or application-specific integrations, presenting a scalable model for embedding real-time artificial intelligence into everyday computing environments.

Author 1: Divij H. Patel

Keywords: Keystroke dynamics; keystroke simulation; AI-Assisted Interfaces; open-source AI; key-chord recognition; API integration; real-time interaction systems

Paper 6: Best Practices to Train Accurate Deep Learning Models: A General Methodology

Abstract: In recent years, the field of computer science has experienced great changes due to the incredible advances in the field of artificial intelligence. Deep Learning models are responsible for most of them since the biggest milestones occurred in 2012 when AlexNet won the image classification challenge called ImageNet. These models have demonstrated great performance in different types of complex tasks like image restoration, medical diagnosis, or object recognition. Their disadvantages are related to their high data dependency, which forces experts in the field to follow a precise methodology to obtain accurate models. In this study, we describe a complete workflow that begins with the management of the raw data until the in-depth interpretation of the performance of the models. This should be taken as a high-level consultation document describing good practices that should be applied. Apart from the step-by-step methodology, we present different use cases that correspond to the two main problems of the field: classification and regression.

Author 1: Alberto Nogales

Author 2: Ana M. Maitín

Author 3: Álvaro J. García-Tejedor

Keywords: Methodology; artificial intelligence; deep learning

Paper 7: A Constraint-Driven Conversational Architecture for Staged Task Interaction Using Large Language Models

Abstract: Large language models (LLMs) enable flexible conversational interfaces, but remain difficult to deploy in structured, staged tasks that require controlled progression, bounded information disclosure, and task validity. Prompt-based control is inherently probabilistic and has been shown to degrade under multi-turn interaction, leading to premature solution disclosure, stage skipping, loss of state coherence, and constraint violations. This study presents a constraint-driven conversational architecture that separates probabilistic language generation from deterministic task governance. An external control layer manages dialogue state, stage transitions, and constraint enforcement, while a simulation layer represents task logic independently of the LLM. We instantiate the architecture in the context of customer discovery tasks to illustrate how staged processes and bounded disclosure can be operationalized without embedding task logic directly into prompts. This work focuses on architectural design and control mechanisms rather than outcome evaluation, offering a reusable architectural pattern for LLM-driven conversational systems that must preserve staged progression, enforce constraints, and prevent premature disclosure during multi-turn interaction.

Author 1: Joseph Benjamin Ilagan

Author 2: Jose Ramon Ilagan

Keywords: Large language models; conversational systems; task-oriented dialogue; deterministic control; constraint enforcement; staged interaction

Paper 8: Measurement of Dataset Quality and Computational Quality of Graph Neural Networks on Analog Integrated Circuit Recognition System

Abstract: To meet the needs of AI design in analog IC design automation and especially in big data applications, the application of sensitive and precise Graph Neural Network (GNN) recognition will become increasingly necessary. This sensitive GNN recognition is needed to accommodate the highly stringent trade-offs on analog IC design requirements, combined with the increasingly large learning data requirements. Furthermore, more precise GNN recognition will be a fundamental requirement for a more sensitive system to accommodate noise in floating point (FP) calculations, namely inaccuracies and imprecision in IEEE 754 FP calculations. In this study, by refining a previously proposed method that uses the output vector representation (OVR) of the untrained Graph Neural Network (GNN), and by exploiting the numerical reproducibility error in FP calculations, a method for measuring the sensitivity and precision of GNNs for use in analog IC design recognition has been proposed. The results of the sensitivity and precision measurements of the proposed method show a complex combination of effects between dataset quality (the presence or absence of data duplication), the sensitivity of the GNN to distinguish each feature in the dataset, the complexity of the calculations that occur in the GNN, and the level of quality of the FP calculations performed by the processing unit related to the non-ideal FP calculations. With certain GNN configurations, the proposed method also succeeded in measuring the difference in FP calculation quality of Central Processing Unit (CPU) and Graphics Processing Unit (GPU), where, for big data applications, from the tests carried out, the maximum amount of data that can be distinguished by the CPU is 25 to 100 times more than with the GPU. Because it only uses untrained GNN OVR and does not involve any training process in obtaining results, the proposed method is still unable to obtain a correlation with the final value of GNN performance after training is complete. The measurement method using GNN OVR involving the learning process is future work.

Author 1: Arif Abdul Mannan

Author 2: Koichi Tanno

Keywords: Big data; Graph Neural Network; Artificial Intelligence; calculation quality measurement; analog circuit design

Paper 9: LayCoder: UI Layout Completion with an Encoder-Only Transformer and Layout Tokenizer

Abstract: The growing complexity of user interface (UI) design calls for effective methods to understand, complete, and refine layout structures. While prior work has focused pre-dominantly on generating UI layouts from scratch, completing partially designed interfaces is equally critical, particularly in iterative design workflows and scenarios involving incomplete prototypes. In this study, we address the UI layout completion task for mobile app screens using an encoder-only transformer architecture with masked modeling and a layout tokenizer. By representing UI elements as discrete tokens, we formulate layout completion as a sequence prediction problem that leverages global context to infer missing components. We evaluate our approach on subsets of the RICO dataset designed with varying constraints on UI element types and spatial overlap, and report results using standard layout metrics: Coverage, Intersection over Union (IoU), Max IoU, and Alignment. The experiments demonstrate that the proposed method achieves substantial improvements over LayoutFormer++, a widely adopted baseline in UI layout generation, particularly in Coverage, and in several cases, IoU, Max IoU, and Alignment. Additional experiments on noise-reduced subsets reveal that dataset curation can enhance spatial consistency but may also reduce Coverage, reflecting an inherent trade-off between completeness and structural precision. These findings highlight both the effectiveness and the limitations of encoder-only masked modeling for layout completion, and underscore the importance of balancing model design with dataset construction when tackling complex UI design tasks.

Author 1: Iskandar Salama

Author 2: Luiz Henrique Mormille

Author 3: Masayasu Atsumi

Keywords: Layout completion; deep learning; encoder-only transformers; masked language modeling; tokenization

Paper 10: Reducing Interference in Human–Robot Collaboration: A Distance-Based Policy for Disassembly Collaborative Task

Abstract: This study investigates human-robot collaborative (HRC) disassembly tasks, focusing on the effects of constrained and unconstrained safety distances on human behavior and movement interference. Separate and shared target configurations were compared using a screw-picking task as a prototype for electronic waste (e-waste) disassembly. The system employ a vision-based approach to track screws, the human hand, and the robot end effector. Performance was evaluated using objective metrics, including completion time, warning distance, and movement heatmaps, as well as subjective workload assessed via NASA-TLX. Results show that the safety algorithm significantly improved movement efficiency by minimizing the spatial distribution compared to the unconstrained trials. Furthermore, significant results from the factorial design showed that separate-target task enhanced human awareness, enabling participants to anticipate warning distances more effectively than the shared-target task. Synchronizing task assignments with safety algorithms reduces the need for human intervention.

Author 1: Wendy Cahya Kurniawan

Author 2: Wen Liang Yeoh

Author 3: Osamu Fukuda

Keywords: Human-robot collaboration; disassembly collaborative task; UI feedback; distance policy; vision-based sensor

Paper 11: Understanding Authentication and Authorization: A Comparative Analysis of Role-Based Access Control (RBAC), Attribute-Based Access Control (ABAC), and Relationship-Based Access Control (ReBAC) Authorization Models

Abstract: This study elucidates the distinctions between authentication and authorization within information security, two fundamental yet frequently conflated concepts. While authentication serves to confirm an entity’s identity, authorization determines the permissible actions that entity may execute. A thorough understanding of these mechanisms is critical for architecting secure, scalable systems and reducing vulnerabilities. The study further explores three widely adopted authorization paradigms using a gymnasium analogy: Role-Based Access Control (RBAC), which assigns privileges based on predefined roles; Attribute-Based Access Control (ABAC), which leverages a dynamic evaluation of user and contextual attributes; and Relationship-Based Access Control (ReBAC), which determines access based on defined relationships among entities. The concluding discussion emphasizes that optimal security is realized when authentication and authorization function cohesively.

Author 1: Madhuri Margam

Keywords: Authentication; authorization; access control; Role-Based Access Control (RBAC); Attribute-Based Access Control (ABAC); Relationship-Based Access Control (ReBAC); information security; least privilege; Zero Trust

Paper 12: Construction of an International Trade Financial Risk Assessment and Prediction Model Based on Big Data Analysis

Abstract: Background: International trade promotes economic growth across nations, while imposing financial risks of currency fluctuations, credit defaults, and market volatility. Although conventional methods of risk evaluation have served well in the past, they are, however, unable to provide risky international trade answers under the dynamic conditions at present. Objective: This study aims to develop data-driven risk assessment and prediction models for financial risks in international trade, with an emphasis on the China trade regime dominated by finance. The purpose is to maximize prediction accuracy and to provide pragmatic risk management solutions. Methods: The study proposed a hybrid method capable of characterizing complex nonlinear correlations by a DNN and, subsequently, estimating prediction outputs with an LR model for enhanced interpretability. Training models on the International Trade and Finance Dataset are augmented by macroeconomic indicators; preprocessing is performed via statistical imputation, feature normalization, and one-hot encoding. Results: With values of 0.9670, 0.0408, 0.0322, and 0.0017 awarded to R², RMSE, MAE, and MSE, respectively, the model stands out as the most capable and accurate in measuring financial risk. However, this hybrid model marries complex features with interpretable features, thereby paving the way for an exquisite instrument for risk assessment. Conclusion & Implications: This study aims to develop a solid framework for predicting financial risks in international trade that will aid financial institutions in decision-making and in developing policies. The findings may be applied to ongoing financial stability assessments for trade risk management.

Author 1: Zeyu Liu

Keywords: International trade; financial risk; big data; predictive modeling; China’s economy

Paper 13: Attention and Representation Learning in Byte-Level Digital Forensics: A Survey of Methods, Challenges, and Applications

Abstract: Byte-level analysis has become an essential capability in digital forensics, enabling content-based investigation when file system metadata, headers, or structural information are unavailable or unreliable. Recent advances in deep learning allow forensic systems to learn discriminative features directly from raw byte streams; however, the growing diversity of representation strategies, architectural designs, and attention mechanisms makes it difficult to assess their relative effectiveness and practical suitability. This study presents a structured survey of representation learning and attention-based approaches for byte-level digital forensic analysis. We examine statistical, embedding-based, image-based, sequential, and hybrid representations, and analyze how architectural choices and attention mechanisms influence performance, robustness, and scalability. Across the literature, hybrid representations combined with lightweight convolutional backbones and selective attention mechanisms consistently provide a favorable balance between accuracy and computational efficiency. The survey also reviews key forensic applications, including file fragment classification, malware and binary analysis, network payload forensics, and encrypted or compressed data triage. In addition, we critically discuss challenges related to distribution shift, dataset bias, adversarial vulnerability, interpretability, and reproducibility, along with practical considerations for deployment in large-scale forensic pipelines. By synthesizing architectural trends, operational constraints, and reliability concerns, this work identifies critical research gaps and provides a structured foundation for the development of robust and trustworthy byte-level forensic learning systems.

Author 1: Teena Mary

Author 2: Sreeja CS

Keywords: Byte-level digital forensics; representation learning; attention mechanisms; file fragment classification; deep learning; forensic robustness

Paper 14: Estimation of Landslide Hazard Zones Using Deep Learning Based on Diverse Geospatial Data

Abstract: Traditional landslide hazard mapping in Japan relies on labor-intensive field surveys, which are slow, costly, and fail to update dynamically amid rising climate-driven disasters like the 2018 Heavy Rain Event, leaving gaps in timely evacuations. This study addresses these challenges by proposing a semantic segmentation framework using ResUNet to fuse Sentinel-2 optical, Sentinel-1 SAR amplitude, DEM-derived Terrain Ruggedness Index (TRI), and JAXA land cover data, tackling class imbalance with BCE + Dice loss and providing probability/uncertainty maps via 4-TTA for robust hazard delineation under adverse weather. The principal aim is to enable operational, weather-robust hazard zone extraction with AUC upto 0.89 (best multimodal configuration), outperforming single-modality baselines (e.g., optical-only AUC 0.74; SAR-only 0.69) through synergistic feature fusion, while highlighting multimodal SAR's edge for cloud-obscured scenarios. Validated on Hiroshima Prefecture data—Japan's highest-risk region with ~32,000 hazard spots—this approach demonstrates pre/post-disaster change detection, but reveals limitations in spatial generalization due to region-specific training.

Author 1: Kohei Arai

Author 2: Kengo Oiwane

Author 3: Hiroshi Okumura

Keywords: SAR; Optical; ResUNet++; land cover classification; Digital Elevation Model; landslide hazard zones; The Heavy Rain Event of July 2018

Paper 15: A Hybrid Framework Integrating GNN-LSTM-CNN to Map the Impact of MSME User Behavior on Digital Transformation

Abstract: MSMEs need a recommendation system that simultaneously captures the evolution of user intent over time and the relationship structure between entities (users, products, sessions, categories, and security events). The problem with this research is that LSTM excels in sequence, but its performance drops on a rare timeline, a general situation in MSME logs. GNN is strong for cross-entity relationships, but it does not explicitly model temporal dynamics. The gap arises because many pipelines still separate design signals, temporal behavior, and session security, reducing explainability and long-term reliability. Our contribution proposes a calibrated and security-aware hybrid that integrates CNN, heterogeneous GNN with reverse edges, and LSTM for behavioral sequences. Multitask-trained models (BCE for purchase links and λ· BCE for session risk) with L2 regularization and post-practice calibration, chronological data sharing prevents leakage. The goal is to design and evaluate CNN-plus CNN-enhanced GNN-LSTM hybrids to improve the accuracy of recommendations and reduce risk. The results on partner MSME data: ROC-AUC 0.965 (val)/0.946 (test), PR-AP 0.943/0.910; risk ROC-AUC 0.984, PR-AP 0.982, surpassing a CNN-BiLSTM baseline (0.93/0.91). Brier scores 0.161 (links) and 0.176 (risk) enable safer personalization. Going forward, we are focusing on per-segment calibration with ECE/MCE reporting, compute efficiency, multimodal expansion, ablation, and explainability (GNNExplainer, CNN saliency), as well as online retraining and drift monitoring to maintain production performance.

Author 1: Jani Kusanti

Author 2: Erni Widiastuti

Author 3: Bintara Sura Priambada

Author 4: Ramadhian Agus Triono Sudalyo

Author 5: Masdava Aviv Masyayissa

Author 6: Rizki Adhi Pratama

Keywords: CNN; e-commerce; hybrid model; LSTM; MSME; user behavior prediction

Paper 16: A Computer Vision–Based Method for Determining the Vaccine Injection Position in Pangasius Fingerlings

Abstract: Pangasius farming in the Mekong Delta is a major component of Vietnam’s aquaculture industry, characterized by large-scale production, intensive farming practices, and significant contributions to export revenue. However, vaccination of Pangasius fingerlings is still predominantly performed manually, resulting in low productivity, high labor demand, and inconsistent injection accuracy, which limit large-scale deployment in commercial hatcheries. Therefore, there is an urgent need to develop an automated and accurate vaccination method for Pangasius fingerlings. This study proposes a novel computer vision–based approach for non-contact measurement of Pangasius fingerlings and accurate determination of the vaccine injection position. The proposed method leverages the image processing capabilities of the OpenCV library in combination with statistical morphological characteristics of Pangasius fingerlings to localize the injection position. The Python-implemented algorithm is lightweight and can run on an embedded Raspberry Pi platform, supporting practical in-field deployment. Experimental results demonstrate an average positioning accuracy of 97.65%, confirming the effectiveness of the proposed approach and its potential to serve as a technological foundation for automated vaccination systems in Pangasius aquaculture.

Author 1: Nguyen Phuc Truong

Author 2: Luong Vinh Quoc Danh

Author 3: Nguyen Chanh Nghiem

Keywords: Computer vision; fish length measurement; fish vaccination; injection position; OpenCV; Pangasius

Paper 17: Exposure-Based Media Mix Modeling Using Machine Learning and Genetic Algorithms

Abstract: Media budget allocation remains a persistent challenge in the advertising industry. Inefficient spending and biased planning decisions often reduce campaign effectiveness. Advertisers struggle to balance investments across television, radio, press, and digital platforms while managing diminishing returns. This study proposes a data-driven media mix determination model that integrates supervised machine learning with genetic algorithm-based optimization. The objective is to maximize audience reach while maintaining cost efficiency. Unlike traditional media mix models that rely on aggregated medium-level performance, this study adopts an audience exposure-based modelling approach. Facebook and YouTube are used as digital media platforms in this study. Television and digital models are trained using exposure-based reach measures, such as 1 plus and 2 plus reach. Machine learning models, including decision trees, random forests, XGBoost, and LightGBM, are evaluated to capture complex and nonlinear relationships between spend and exposure-based reach. Smoothed reach response curves are used to identify efficiency levels and saturation points for each medium. A genetic algorithm is then applied to derive the optimal budget allocation across media under efficiency, reach, and cost constraints. The model is trained using real advertising data from the Sri Lankan market, ensuring practical relevance and applicability. Although the analysis is based on a country-specific dataset, the model is transferable to markets of similar scale. This study contributes to the literature by introducing an exposure-driven media mix modelling approach that improves media budget planning accuracy and supports more effective advertising decision-making.

Author 1: Thejan Dulara

Author 2: Indra Mahakalanda

Author 3: Prasanga Jayathunga

Keywords: Media mix determination; saturation points; audience reach; audience exposure

Paper 18: An Ensemble Learning Framework with Metaheuristic Optimization for Credit Card Fraud Detection

Abstract: Credit card fraud detection is a major challenge in the financial system due to the characteristics of highly unbalanced data. This study proposes an ensemble learning approach combined with hyperparameter optimization using a Genetic Algorithm to improve the performance of fraud transaction detection. The results of the experiment showed that Random Forest achieved the best performance with a perfect Recall of 1.00 and an F1-Score of 0.903, outperforming the Stacking and Bagging models. Although the optimization significantly increases the training time, this method manages to accelerate the inference time to 0.0290 seconds, making it very feasible to apply to real-time banking security systems that require instant validation. This study confirms the effectiveness of integrating ensemble learning and metaheuristic optimization in dealing with the problem of unbalanced data.

Author 1: Agung Nugroho

Author 2: Muhtajuddin Danny

Author 3: Ismasari Nawangsih

Keywords: Fraud detection; ensemble learning; Genetic Algorithm; Random Forest; real-time detection

Paper 19: A Lossless Medical Image Compression Framework Using Logic Minimization

Abstract: Medical image compression is an active research area owing to the growth of the volume of medical image data in digital form. A method for lossless compression of medical images was proposed using logic minimization. The grayscale medical image is split into bit planes, and each bit plane is divided into blocks of fixed size, e.g., 8 × 4. The binary bit stream resulting from each of these blocks is treated as the output of a Boolean function, and logic minimization is attempted for a compact form. If this step fails to give a compact representation, the bits are stored as such. In this study, an extendable framework for lossless compression of medical images is presented. The bit plane is adaptively divided using a Quadtree to capture large uniform areas as leaf nodes. The non-uniform blocks at the leaf node are subjected to a logic minimization approach, as in the case of fixed-size blocks in previous related work. To improve the result further, the original image is gray-coded as a pre-processing step. On the bit plane, an XOR operation is done for the current block with the neighboring block for redundancy removal. This framework allows further exploration with the incorporation of other Boolean function representation techniques to enhance the compression.

Author 1: Swathi Pai M

Author 2: Jacob Augustine

Author 3: Pamela Vinitha Eric

Keywords: Medical image compression; lossless compression; Quadtree; logic minimization; bit plane encoding; XOR operation; gray coding

Paper 20: An Empirical Evaluation of Multivariate Temporal Convolutional Networks with Global Market Indicators for Forecasting the Indonesian LQ45 Index

Abstract: This study develops a multivariate Temporal Convolutional Network (TCN) framework to forecast the LQ45 stock index using daily time-series data from January 2015 to January 2025. The objective is to examine whether incorporating global market indicators, namely the Volatility Index (VIX), Brent crude oil price, and the Effective Federal Funds Rate (EFFR), alongside lagged LQ45 values, improves forecasting performance in Indonesia’s equity market. Two comparison models are considered: an Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX) model as a statistical baseline and a univariate TCN as a deep learning benchmark. Data preprocessing includes normalization and a seven-day sliding-window framing. Forecasting accuracy is evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), complemented by validation and interpretability analyses. The results show that while the multivariate TCN captures interactions among multiple temporal features, it does not provide measurable performance advantages over ARIMAX or the univariate TCN. The lagged LQ45 series exhibits the strongest predictive contribution, followed by EFFR with a stable secondary effect, whereas Brent oil prices and VIX display weak and unstable influences. These findings suggest that in short-horizon forecasting of relatively stable emerging markets, integrating exogenous variables showed no performance improvement and was associated with higher forecast error when their predictive structure was unstable, highlighting the trade-off between feature complexity and model robustness.

Author 1: Yohanes Marakub

Author 2: Muhammad Zarlis

Keywords: Deep learning; financial forecasting; multivariate forecasting; TCN; time series analysis

Paper 21: Integrating ABM and GIS for Flood Evacuation Planning: A Systematic Review and Future Direction

Abstract: This systematic review examines the integration of agent-based modeling (ABM) and Geographic Information Systems (GIS) in flood evacuation planning from 2015 through early 2025. This review aims to systematically evaluate how ABM and GIS have been integrated in flood evacuation research, identify methodological gaps, and propose a structured framework to guide future model development. Using PRISMA guidelines, 67 studies were selected and analyzed to uncover methodological trends, empirical gaps, and policy relevance in this growing research domain. Using the PRISMA 2020 framework, the analysis reveals a dominant reliance on mesoscopic modeling (43%), limited real-time data integration (17.9%), weak empirical validation practices (16.4%), and minimal machine learning adoption (4.5%). To structure the evolving landscape, a conceptual integration framework is proposed to classify studies by modeling scale, data fidelity, and validation strategy. This framework highlights a gradual shift toward behaviorally realistic, spatially precise, and policy-relevant evacuation models. Persistent challenges include limited validation practices, weak real-time responsiveness, and insufficient policy integration. Conclusions were drawn by identifying five research priorities: AI integration, real-time enhancement, multi-hazard modeling, empirical grounding, and participatory policy co-design. This review offers actionable insights for advancing robust, scalable, and operational ABM-GIS systems in disaster risk reduction.

Author 1: Kabir Musa Ibrahim

Author 2: Abubakar Ahmad

Author 3: Noor Akma Abu Bakar

Author 4: Mazlina Abdul Majid

Author 5: Azamuddin Rahman

Keywords: Agent-based modeling; GIS integration; flood simulation; spatial modeling; evacuation dynamics; multi-agent systems

Paper 22: A Novel Lightweight Explainable Multilayer Adaptive RNN-Based Intrusion Detection Framework

Abstract: A rapid increase in the instances of cyberattacks has been observed with the expanding digitization. This leads to an urgent and critical need for developing robust intrusion detection systems (IDS) which can identify the occurrences of malicious activities within the network traffic. The present work proposes a novel, explainable, multilayer, lightweight adaptive IDS based on a Recurrent Neural Network (RNN). The purpose of this proposed IDS is to improve threat detection capabilities, especially low frequency high severe attack. The performance of the proposed IDS is evaluated using the UNR-IDD dataset. The network traffic is classified into normal and attack categories to assess the effectiveness of the proposed IDS. Two separate IDS models are developed. Model A is used to detect attacks on the basis of the frequency of the attacks, and Model B detects threats based on the severity of the attacks. Through the layered approach, the overall detection accuracy of 95.7% is achieved in Model A, and 97.5% is obtained in Model B. The present work highlights that the proposed IDS shows a remarkable improvement in the detection of less frequent severe attacks in comparison to existing IDS. The comparative result analysis of RNN-based IDS with Machine Learning models such as LR, Naïve Bayes, Cat Boost, Random Forest and Multilayer Perceptron models shows RNN-based IDS has outperformed the Machine Learning models. Explainable AI (XAI), the SHAP method is used for better interpretation of the proposed decisions. XAI helps to identify the network traffic that can influence predictions and detect potential biases. It also helps researchers and practitioners to validate model behaviour and establish trust in the system’s outputs.

Author 1: Nidhi Srivastav

Author 2: Rajiv Singh

Keywords: IDS; network security; adaptive techniques; RNN; cybersecurity; explainable AI; UNR-IDD

Paper 23: Empirical Validation of Learnability Factors in Web-Based AR: Insights from the LEMARK–Hafsa Model Grounded in Kolb’s Experiential Learning Theory

Abstract: Augmented Reality in higher education is transforming learning by providing immersive environments that enhance cognitive and motivational engagement. Despite growing interest, there remain limited empirically validated learnability factors that can support future instructional models, such as the LEMARK-Hafsa model. This research attempts to bridge the identified gap through statistically validating seven key factors—Motivation, Confidence, Enhanced Focus, Visualization of Invisible Concepts, Satisfaction, Better Lab Experience, and Better Learning—within the LEMARK-Hafsa model grounded in Kolb’s Experiential Learning Theory. Data collected from 291 participants underwent expert validation, data cleaning, exploratory factor analysis, and regression analysis. The exploratory factor analysis confirmed structural validity, with factor loadings ranging from 0.430 to 0.822. The Kaiser-Meyer-Olkin value was 0.769, and Bartlett’s test was significant (p < 0.001), indicating that the data were suitable for factor analysis and supported multiple distinct factors. The regression results showed that Visualization of Invisible Concepts had a statistically significant positive effect on learning outcomes (the normalized regression weight recorded as 0.155, p = 0.031), while Enhanced Focus (p value of 0.091) and Satisfaction (p value of 0.089) were close to significance. Motivation, Confidence, and Better Lab Experience also showed positive, though not statistically significant, effects that were consistent with theoretical expectations. These findings provide empirical support for the statistical adequacy of the proposed LEMARK–Hafsa factors, establishing a validated measurement basis for subsequent theoretical integration and model-level investigation in research on web-based Augmented Reality learning environments in higher education.

Author 1: Sayera Hafsa

Author 2: Mazlina Abdul Majid

Author 3: Shafiq Ur Rehman

Keywords: Experiential learning theory; educational technology; predictive validity; structural validity; AR-based learning; factor validation; LEMARK–Hafsa model

Paper 24: Automating Computation Independent Model Elicitation in MDA using Task-Oriented Dialogue with In-Context Learning

Abstract: The Computation Independent Model (CIM) is a cornerstone of the Object Management Group's (OMG) Model-Driven Architecture (MDA), capturing business requirements and domain knowledge independent of specific technologies. However, the elicitation of CIM requirements is often a manual, time-consuming, and error-prone process, susceptible to ambiguities inherent in natural language. Traditional Natural Language Understanding (NLU) approaches, particularly intent-based systems, exhibit limitations in scalability, contextual understanding, and handling the nuanced, evolving nature of complex requirements. This study proposes a novel approach that integrates Task-Oriented Dialogue (TOD) systems with the In-Context Learning (ICL) capabilities of Large Language Models (LLMs) to automate and enhance CIM requirements elicitation. The proposed framework features a conversational agent that guides stakeholders through structured dialogue flows, translating their natural language inputs into a formal CIM-Domain Specific Language (CIM-DSL). These DSL commands are then transformed into CIM artifacts, such as Business Process Model and Notation (BPMN) diagrams and Unified Modeling Language (UML) use cases. The approach emphasizes quality assurance through interactive validation, consistency checks, and strategies to mitigate LLM limitations. We anticipate this method will significantly improve the accuracy, completeness, and efficiency of CIM construction, thereby strengthening the foundation of the MDA lifecycle.

Author 1: Mohamed EL Ayadi

Author 2: Yassine Rhazali

Author 3: Mohammed Lahmer

Keywords: MDA; CIM; Task-Oriented Dialogue (TOD); In-Context Learning (ICL); Large Language Models (LLMs); requirements elicitation; Domain-Specific Language (DSL); Artificial Intelligence (AI); Natural Language Understanding (NLU); BPMN

Paper 25: Mapping Research on Artificial Intelligence in Customer Experience: A Bibliometric Analysis

Abstract: The purpose of this article is to conduct a comprehensive bibliometric analysis of research on artificial intelligence and customer experience. The study data was extracted from Scopus and Web of Science, focusing on articles published between 2010 and 2025. VOSviewer and Biblioshiny software were used to map the intellectual landscape of the interaction between artificial intelligence and customer experience, identifying growth in scientific output, geographical distribution and collaboration, influential publications, leading authors, word co-occurrence, leading journals, and thematic trends. The study reveals that research only began in 2017, then interest in this topic began to grow, but most publications date from 2025. Furthermore, the findings reveal that research is concentrated in a limited number of leading countries which are more advanced in terms of technological infrastructures. This field of research is gradually evolving towards greater specialization and increased use of AI to improve customer experience in terms of personalization, decision-making and automation of service.

Author 1: Firdaws Hayoun

Author 2: Brahim Ouabouch

Author 3: Youssef Aatif

Author 4: Taoufiq Yahyaoui

Author 5: Fatima Zahra El Arbaoui

Keywords: Artificial intelligence; customer experience; marketing; personalization

Paper 26: AI-Driven Robotic Waste Sorting for Techno-Economic Assessment in Urban Indonesia

Abstract: Urban centers in Indonesia are facing increasing pressure in managing municipal solid waste as a result of rapid population growth, rising labor costs, and stricter demands for high-purity recyclable materials. Manual sorting at Material Recovery Facilities has become progressively less efficient and economically burdensome under these conditions. This study presents an artificial intelligence-driven robotic waste sorting system designed and evaluated under real operational conditions in Jakarta and South Tangerang. The system integrates YOLO based object detection, vision-guided robotic manipulation, real-time processing hardware, a multi-axis gantry system with stepper motors, and a custom conveyor mechanism to deliver waste items to the sorting cell. Unlike previous studies that mainly focus on algorithmic accuracy or laboratory-scale validation, this work combines real-world technical performance assessment with a localized techno-economic analysis. Experimental results show an average sorting accuracy of 90%, a material purity of 95.1%, and a throughput of 50 items per minute, outperforming typical manual sorting performance. An economic evaluation based on local wage levels, electricity tariffs, and recyclable market prices indicates a payback period of 4.3 to 4.9 years. The main contributions of this study lie in integrating AI vision and robotic sorting into unstructured urban waste environments, in empirical validation under Indonesian operating conditions, and in demonstrating economic feasibility for emerging economies. Although the case study focuses on Jakarta and South Tangerang, the findings are relevant for metropolitan areas across the Global South seeking more efficient and sustainable waste management solutions.

Author 1: Ida Nurhaida

Author 2: Mohammad Nasucha

Author 3: Hari Nugraha

Keywords: Smart waste sorting; AI vision; YOLOv12; computer vision; material recovery facility; techno-economic assessment; urban Indonesia; circular economy

Paper 27: A Unified Benchmarking Framework for Offline Handwritten Signature Verification Using Deep Learning Architectures

Abstract: Offline handwritten signature verification (OSV) remains a challenging biometric task owing to the subtle variability of genuine signatures and the sophistication of skilled forgeries. This study introduces a unified benchmarking framework for evaluating eight deep learning architectures—CNN Shallow, CNN Deep, ResNet 18, ResNet 34, MobileNetV2, EfficientNet B0, ViT Tiny, and a CNN–Transformer Hybrid—within a writer-independent Siamese contrastive-learning paradigm. The framework standardizes preprocessing, balanced pair generation, NVIDIA A100 GPU training, and a comprehensive evaluation suite that includes ROC and Precision-Recall curves, Equal Error Rate (EER), calibration analysis, threshold-sensitivity metrics, and embedding visualizations using PCA, t SNE, and UMAP. Experiments conducted on an NVIDIA A100 GPU reveal a clear performance stratification: six architectures achieve perfect verification performance (Accuracy = 1.0, ROC AUC = 1.0, PR AUC = 1.0, EER = 0.0), supported by consistently well separated embedding manifolds and highly stable calibration behavior. In contrast, MobileNetV2 and EfficientNet B0 exhibited elevated EER values and overlapping embeddings, underscoring the limitations of lightweight and compound-scaled models in capturing the fine-grained stroke morphology. The proposed framework establishes a transparent and extensible foundation for future research, enabling fair cross-model comparisons and guiding the development of robust and deployment-ready biometric verification systems. In addition, this study provides the first fully controlled, architecture-agnostic comparison of CNNs, residual networks, lightweight mobile models, and transformer-based architectures under identical, writer-independent conditions. By eliminating variability in preprocessing, pair generation, and training configuration, the framework isolates the true effect of the architectural design on the verification performance. The findings highlight the importance of embedding separability, calibration stability, and threshold robustness—factors often overlooked in prior OSV research but essential for real-world deployment.

Author 1: Eissa Alreshidi

Keywords: Offline handwritten signature verification; Siamese networks; contrastive-learning; deep learning architectures; ResNet; MobileNetV2; EfficientNet; vision transformers; hybrid CNN-transformer models; biometric authentication; forgery detection; embedding separability; writer-independent verification

Paper 28: Deep Learning Based Detection of Prostate Cancer in MRI Using Biopsy-Confirmed Ground Truth

Abstract: Prostate cancer is one of the most common malignancies in men, and accurate lesion segmentation in magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, and disease monitoring. Manual delineation by radiologists is time-consuming and subject to interobserver variability. This study presents an automated, deep learning-based framework for 3D prostate lesion detection using modified U-Net architectures, guided by pathology-informed ground truth. The proposed approach leverages biopsy-verified lesion masks derived from the PROSTATEx and PROSTATEx2 datasets, ensuring biologically validated reference labels. Method 1 uses dice loss optimization to train a simplified 3D U-Net on full volume MRI data, while Method 2 uses a patch-based 3D U-Net with advanced preprocessing, extensive data augmentation, and a dice focal loss to reduce class imbalance and improve lesion localization. With a Dice similarity coefficient (DSC) of 92.3% and an intersection over union (IoU) of 87.8%, the quantitative data shows that the patch-oriented network performs better in segmentation. In contrast to models trained only on radiologist annotations, the work shows that pathology-informed learning improves lesion delineation accuracy, highlighting its potential for strong clinical translation in MRI-guided prostate cancer detection.

Author 1: Samana Jafri

Author 2: Gajanan Birajdar

Keywords: Prostate lesion; 3D U-Net; MRI; biopsy confirmed lesion masks

Paper 29: A Composite Approach Extracting Software Quality Attributes Assessing Operational Profiles

Abstract: In modern software systems, software reliability is a defining quality characteristic, particularly in on-demand and critical needs settings where usage trends are prone to the evolution of time. The traditional black box reliability models do not sufficiently capture these dynamics, especially in component-based and reusable software architectures. In a bid to overcome such limitations, the present study presents the Level-Wise Composite Software Reliability method based on Operational Profile Evaluation (LCSR-OPE). The suggested model combines package-level measurements of software, software operational profile modelling and probabilistic analysis to approximate and refine software dependability in real usage contexts. The construction of user profiles through structural and complexity metrics clustering is carried out, but the operational behavior is measured by probability density functions in order to concentrate on testing and fault detection. Moreover, defect identification and reliability testing are narrowed down with a machine-learning-based fault classification model. Experimental performance on NASA software fault data sets establishes that the suggested method achieves high levels of fault-detection and reliability estimation as compared to traditional reliability models. The findings verify the effectiveness of the use of operational profiles and the level-based composite analysis in the reliability engineering of software excellence.

Author 1: Sudhakar Kambhampati

Author 2: G Vamsi Krishna

Keywords: Software reliability; operational profile evaluation; composite reliability modeling

Paper 30: Robust Medical Image Reconstruction Using a Self-Evolving Encoder–Decoder and Adaptive Convolutional Power Scaling

Abstract: Robust medical image reconstruction is a critical requirement for accurate diagnosis and clinical decision-making, particularly when images are affected by degradation, noise, or low resolution. Conventional encoder–decoder-based reconstruction methods compress input images into low-dimensional representations and subsequently decode them into high-resolution outputs; however, such approaches often suffer from artifacts and loss of fine anatomical details under severe degradation. To address these limitations, this work proposes a robust medical image reconstruction framework using a self-evolving encoder–decoder and adaptive convolutional power scaling. The proposed super-resolution model incorporates a dynamic encoder and decoder that adaptively evolve during training to capture color contrast, structural similarity, and high-frequency details from medical images. An MLP enhanced with an adaptive power flex layer is embedded within the reconstruction pipeline, enabling learnable power-based feature scaling through weight-wise modulation and initialization. This mechanism improves feature discrimination and stabilizes the reconstruction of subtle anatomical structures. The DRIVE and CHASE_DB1 retinal image datasets are employed for experimental validation, with appropriate preprocessing applied before training and testing. The selected images are processed through the proposed super-resolution model, and performance is quantitatively evaluated using PSNR, SSIM, sensitivity, and specificity metrics. Experimental results demonstrate that the proposed method achieves significant improvements in reconstruction quality and robustness compared to existing approaches, yielding enhanced perceptual quality and structural fidelity in reconstructed medical images. These findings indicate that the proposed self-evolving encoder–decoder with adaptive convolutional power scaling is well-suited for reliable medical image reconstruction applications.

Author 1: Dhanusha P B

Author 2: J. Bennilo Fernandes

Author 3: A. Muthukumar

Author 4: A. Lakshmi

Keywords: Dynamic encoder and decoder; power flex model layer; high resolution images; weight initialization; adaptive convolutional power scaling

Paper 31: Cybersecurity Awareness Among Undergraduate Students in Saudi Arabia: A Quantitative Study

Abstract: Cybersecurity awareness has become increasingly important as digital services expand rapidly in Saudi Arabia. This study presents a quantitative assessment of cybersecurity awareness among undergraduate students in Riyadh and Jeddah. A structured questionnaire was administered to 177 students to evaluate awareness across three dimensions: internet usage, information security practices, and social media and smartphone security. Statistical analyses, including descriptive statistics, principal component analysis, regression, and cluster analysis, were applied to identify patterns and influencing factors. The results indicate a moderate overall level of cybersecurity awareness, with stronger adoption of passive security measures such as antivirus software and firewalls, and weaker engagement with proactive practices, including password management, VPN usage, and two-factor authentication. A key finding is the identification of a smartphone-related security gap: students who rely exclusively on smartphones show significantly lower awareness. The study highlights the influence of device usage and institutional context on cybersecurity behavior and provides recommendations to enhance cybersecurity education and awareness initiatives in higher education.

Author 1: Turky A. Saderaldin

Author 2: Ali Abuabid

Keywords: Cybersecurity awareness; undergraduate students; smartphone security; Saudi Arabia

Paper 32: An Analytical Study of Data Augmentation Across Audio Representations for Infant Cry Classification

Abstract: Several multidisciplinary studies consider an infant’s cry as a valuable source of information, particularly for parents, caregivers, and medical professionals. From a signal processing viewpoint, infant cries can be represented either in the time domain (one-dimensional or 1D raw waveform) or in the time–frequency domain (two-dimensional or 2D spectrogram-based representations). However, the impact of these representations on classification performance, particularly under constrained and imbalanced dataset conditions, remains insufficiently explored. This study presents a comparative analysis of 1D and 2D convolutional neural networks applied to waveform and spectrogram representations of infant cries. Due to the significant class imbalance of the dataset, we employed data augmentation techniques. Experimental results show that the 1D CNN achieved 95% training accuracy and 91% validation accuracy, indicating a relatively small generalization gap. In contrast, 2D CNN reached 98% training accuracy but remained below 91% on the validation set, revealing a larger gap and suggesting potential overfitting to the augmented data.

Author 1: Meriyem Ghanjaoui

Author 2: Abdelaziz Daaif

Author 3: Abdelmajid Bousselham

Author 4: Sajid Rahim

Author 5: Ahmed Bouatmane

Author 6: Mohamed Elyoussfi

Keywords: CNN; waveform; spectrogram; deep learning; baby cries

Paper 33: Hybrid BERT–BiLSTM Architecture for Enhanced Cyber Threat Intelligence Classification

Abstract: Cyber Threat Intelligence (CTI) plays a crucial role in supporting proactive cybersecurity defence by offering insights into adversarial behaviours and attack tactics. However, CTI data are mainly presented in unstructured natural language, characterised by dense technical terminology, implicit attack semantics, and sequential descriptions of multi-stage threat activities. While transformer-based language models such as BERT have shown strong contextual representation abilities, they are naturally limited in explicitly modelling long-range sequential dependencies that often occur in CTI narratives. On the other hand, recurrent neural networks like BiLSTM effectively capture temporal dependencies, but lack deep contextual understanding. This study proposes a hybrid BERT–BiLSTM architecture that combines the contextual semantic strengths of transformers with the sequential learning abilities of bidirectional recurrent networks for improved CTI text classification. In the proposed framework, BERT acts as a feature extractor to produce contextualised token representations, which are then processed by a BiLSTM layer to model the progression of threats before final classification. A unified experimental setup is used, employing a publicly available CTI dataset, with consistent preprocessing, training strategies, and evaluation metrics to ensure fair assessment. Experimental results show that the proposed hybrid model consistently surpasses standalone BERT and BiLSTM baselines across multiple performance metrics, including accuracy and macro F1-score, with significant improvements especially in minority and semantically ambiguous threat categories. Further analysis indicates that the hybrid architecture effectively reduces common misclassification patterns caused by overlapping attack stages and implicit indicators. These findings demonstrate the effectiveness of combining contextual and sequential modelling approaches for CTI analysis. The proposed BERT–BiLSTM framework provides a robust and interpretable solution for automated CTI classification and offers practical insights for deploying hybrid deep learning architectures in real-world cybersecurity intelligence systems.

Author 1: Syarif Hidayatulloh

Author 2: Salman Topiq

Author 3: Ifani Hariyanti

Author 4: Dwi Sandini

Keywords: Cyber Threat Intelligence; hybrid deep learning; BERT–BiLSTM architecture; text classification; sequential modelling; cybersecurity natural language processing

Paper 34: An Offline Estimation Method for Hip and Knee Joint Angles of Lower Limbs Based on Quaternion and DTW Time Alignment

Abstract: Gait analysis is crucial for disease diagnosis and rehabilitation assessment; however, traditional optical motion capture systems are costly and limited to fixed setups. This study presents an "Offline Estimation Method for Hip and Knee Joint Angles of Lower Limbs Based on Quaternion and DTW Time Alignment." The method uses quaternion fusion of inertial measurement unit (IMU) orientations and employs Sakoe-Chiba constrained Dynamic Time Warping (DTW) to eliminate a 42 ms initial offset and a 150 ms cumulative drift. Combining N-pose calibration with heel velocity event detection allows for the offline calculation of hip and knee joint angles. Data were collected from eight healthy participants during flat walking and stair ascent/descent scenarios, with the Noraxon Ultium Motion system serving as the reference. Results show that DTW reduces the average root mean square error (RMSE) by 29.1%; specifically, "the RMSE for hip flexion reaches 4.1°, while the overall knee joint RMSE is 10.2°, with correlation coefficients ≥0.87. Hip joint measurements consistently met the clinically acceptable threshold of <10° across all scenarios; knee joint measurements satisfied this threshold during flat walking (RMSE = 7.8°) but exceeded it during stair negotiation (RMSE = 11.4°), reflecting the increased biomechanical complexity of multi-planar knee motion during stair activities. This study provides a low-cost, high-precision solution for the post-hoc offline estimation of hip and knee joint angles. The proposed method is specifically designed for retrospective gait data analysis rather than real-time feedback, offering a scalable strategy for early screening of gait abnormalities and clinical assessment in home and community rehabilitation settings.

Author 1: Yuling Zhang

Author 2: Yuejian Hua

Author 3: Shengli Luo

Author 4: Xiaolong Shu

Author 5: Hongyan Tang

Author 6: Hongliu Yu

Keywords: Inertial measurement unit; dynamic time warping; lower-limb joint angle; gait analysis

Paper 35: Radiomics Feature Profiling of Brodmann Regions in Structural MRI: A Machine Learning Study of Intensive Verbal Memorisation (Huffaz vs Controls)

Abstract: Memorisation-based cognitive training has been hypothesized to relate to experience-dependent brain plasticity; however, quantitative evidence at the regional level remains limited. We hypothesized that radiomics descriptors extracted from Brodmann-area volume-of-interest (VOI) regions in pre-processed structural MRI would contain sufficient information to discriminate Quran memorizers (Huffaz) from non-memorizers (controls), and we evaluated this hypothesis using a fully nested validation framework. T1-weighted MRI volumes were pre-processed using a voxel-based morphometry pipeline, and VOIs were defined using Brodmann-area masks. Using PyRadiomics, first-order and texture features were extracted per VOI and combined into a feature matrix for classification. Models were evaluated using repeated nested cross-validation (outer 5-fold × 10 repeats; inner 5-fold for tuning), with ROC-AUC as the primary metric. Random Forest achieved the strongest discrimination (AUC = 0.6704 ± 0.1792), followed by Logistic Regression (AUC = 0.5948 ± 0.2153), while SVM with an RBF kernel underperformed (AUC = 0.4356 ± 0.1927). One-sided testing against chance (AUC = 0.5) indicated above-chance performance for Random Forest and Logistic Regression, but not for SVM-RBF. These results should be interpreted as exploratory because the cohort is small (n = 47) and no independent external validation cohort was available. Practically, the observed effect sizes suggest that VOI-based radiomics may capture detectable group-associated imaging signatures under the current preprocessing and VOI assumptions, motivating validation on larger cohorts, sensitivity analysis (e.g., discretization/normalization settings), and assessment of probability calibration.

Author 1: Mohd Zulfaezal Che Azemin

Author 2: Iqbal Jamaludin

Author 3: Abdul Halim Sapuan

Author 4: Mohd Izzuddin Mohd Tamrin

Keywords: Radiomics; PyRadiomics; Magnetic Resonance Imaging (MRI); voxel-based morphometry (VBM); neuroplasticity; Huffaz; Brodmann-areas; Volume of Interest (VOI); texture analysis; machine learning; nested cross-validation; ROC-AUC; Random Forest

Paper 36: Cognitive Assistance for Prosopagnosia Patients Using Landmark-Based Identity and Emotion Recognition

Abstract: Prosopagnosia is a neurological condition that significantly affects the social interaction and quality of life of individuals. Existing assistive systems mainly focus on either face identity recognition or face emotion recognition, limiting their effectiveness in cognitive assistive scenarios. To overcome these, an integrated framework is needed that jointly addresses identity and emotion recognition to efficiently support prosopagnosia patients, as discussed in this study. The proposed system includes two separate modules: a face identity recognition module and a face emotion recognition module. The proposed system detects and aligns faces using Multi-Task Cascaded Convolutional Networks (MTCNN) with five-point landmark alignment. Face identity recognition is performed using an EfficientNet-B3 backbone to extract 512-dimensional facial embeddings, which are matched against a Structured Query Language (SQL) database using cosine similarity. Then, facial landmarks are detected along with emotion recognition, using the Dlib library, and are structured as a graph for High-order Graph Attention Network (HoGAN)-based relational interactions detection. The proposed system is trained using a joint loss function to efficiently provide real-time assistive feedback. The system achieves high recognition performance, with AUC values of 99.8% and 99.5% in both face identity recognition and face emotion recognition modules.

Author 1: Bhavana Nagaraj

Author 2: Rajanna Muniswamy

Keywords: Cognitive assistance system; EfficientNet-B3; emotion recognition; face identity recognition; prosopagnosia

Paper 37: Comparative Analysis of Neural Network Architectures for Classifying Depressive Content in Social Networks

Abstract: Depression-related language on social media provides measurable signals for population-level mental-health research, yet model selection remains sensitive to evaluation protocol, domain shift, class imbalance, and computational constraints. This study benchmarks CNN, LSTM, and transformer encoders (BERT, RoBERTa, DistilBERT, and MentalBERT) for binary depression-indicative versus control classification on a unified corpus of 19,800 English posts/comments aggregated from three platforms (Reddit, Twitter, and Facebook) under a consistent preprocessing pipeline. We report two complementary evaluation protocols: (1) a fixed-split single-run baseline for a comparable snapshot, and (2) a five-seed repeated-run protocol with statistical testing (effect sizes and multiple-comparison correction) to quantify variability and reduce sensitivity to initialization effects. Under repeated-run reporting, MentalBERT achieves the best overall performance (F1 = 0.918 ± 0.005; AUC = 0.962 ± 0.002), while CNN/LSTM baselines show lower robustness under cross-platform transfer. Cross-domain experiments reveal a consistent performance drop relative to in-domain evaluation, confirming non-trivial platform shift and motivating robustness-aware reporting for deployment-oriented settings. In addition to predictive metrics, we report training time, inference latency, and derived throughput to support practical model selection for use cases such as moderation pipelines and screening/triage dashboards.

Author 1: Yntymak Abdrazakh

Author 2: Rita Ismailova

Author 3: Nurseit Zhunissov

Author 4: Arypzhan Aben

Author 5: Anuarbek Amanov

Author 6: Aigerim Baimakhanova

Keywords: Depression detection; social media text; natural language processing; cross-platform evaluation; robustness; statistical significance testing; transformer-based models; CNN; LSTM; BERT; MentalBERT

Paper 38: Attention-Based Capsule Network with Vision Transformer for Underwater Hyperspectral Image Classification

Abstract: The underwater investigations and research remain challenging due to various underwater distortion factors, scattering, and low wavelength absorption. Hyperspectral imaging helps in obtaining detailed information on each underwater object using the spectral reflectance, using 100 to 300 bands. The Hyperspectral imaging in underwater applications contains the 3D hyperspectral cube, which needs a high level of processing that results in high-accuracy classification. Hence, this study proposes the framework of hybrid deep learning techniques that work on segmentation, feature extraction, and classification processes. The Channel Attention Module (CAM) based U-Net architecture is used for the segmentation process to obtain the spectral spatial characteristics based on the Region of Interest (ROI). The CapsNet Feature extraction helps in obtaining the features of various bands, which helps in the classification of object class-wise using the pose-based relationships. The Vision Transformer (ViT) based classification that depends on the capsule vector token, carries out the multi-class classification by obtaining the global attention among the feature vectors and relationship-based long-range ROI feature data. In this way, the proposed model attains 95.3% accuracy using the maximum IoU of 0.88 and 95.2% of the segmentation process, which helps in achieving 0.99AUC for 8 substrate classes of the underwater HSI dataset.

Author 1: Thiyagarajan B

Author 2: Thenmozhi M

Keywords: Underwater imaging; hyperspectral; U-Net; Vision Transformer; CapsNet; classification; segmentation

Paper 39: Effectiveness of Capsule Networks in Detecting Deepfakes Instead of Traditional CNNs

Abstract: As artificial intelligence has advanced, computer-generated fake content has become increasingly prevalent. Deepfake is an advanced fake creation generated using deep learning-based technologies, and deepfake images, videos, and voices have spread rapidly. The distinction between real and fake content is very hard for the naked eye, making reliable detection essential. Existing deepfake detection methods have achieved success, but still face limitations in keeping pace with the rapid evolution of deepfake generation techniques. In particular, CNN-based approaches may require a large number of parameters and may not fully capture spatial hierarchies. In this research, we investigate whether Capsule Networks can provide an effective and parameter-efficient alternative for deepfake detection. We proposed four different Capsule Network architectures by altering the size, complexity, and configuration. A comparative analysis is conducted against state-of-the-art Capsule models and CNN models across various datasets, utilizing AUC% and the number of parameters as evaluation criteria. For transparency, we note that some baseline CNN and Capsule models follow the training protocols and datasets reported in their original studies, which may differ across implementations. Our experimental results show that the proposed Capsule Network models achieved over 98% AUC% on the evaluated datasets while using fewer parameters than several CNN-based models. These findings suggest that Capsule Networks exhibit greater efficacy in detecting deepfakes compared to traditional CNN-based methods and represent a promising direction for future research.

Author 1: M. C. Weerawardana

Author 2: T. G. I. Fernando

Keywords: Deepfake; deep learning; capsule network; convolutional neural network

Paper 40: Machine-Learning–Assisted Probabilistic Wind Assessment at Sechura, Peru

Abstract: Accurate characterization of wind resources is essential for reliable energy yield estimation and wind farm planning, particularly in regions with limited long-term measurements. This study presents a machine-learning–assisted probabilistic wind assessment in Sechura, Peru, based on multi-year hourly wind data obtained from the NASA POWER database. A representative Typical Meteorological Year (TMY) was constructed to preserve seasonal and diurnal variability while enabling standardized annual energy production (AEP) calculations. Wind speed distributions were modeled using empirical distributions, kernel density estimation (KDE), the Weibull distribution, and Gaussian mixture models (GMM). Statistical evaluation indicates that KDE and GMM reduce the annual RMSE by more than 50% compared to the Weibull model, achieving coefficients of determination above 0.98. Annual energy production is estimated at approximately 1.88 GWh, with differences below 0.3% among probabilistic models. The corresponding capacity factor is approximately 0.25 for a utility-scale wind turbine. The results demonstrate that advanced probabilistic models substantially improve wind speed representation while having a limited impact on integrated annual energy estimates, highlighting the importance of model selection for variability and seasonal analysis rather than for annual yield estimation.

Author 1: Ubaldo Yancachajlla Tito

Author 2: Celso Antonio Sanga Quiroz

Author 3: Edilberto Velarde Coaquira

Author 4: Germán Belizario Quispe

Keywords: Wind resource assessment; probabilistic modeling; machine learning; kernel density estimation; Gaussian mixture model; annual energy production; capacity factor

Paper 41: Ziraai SSI: A Blockchain-Based Self-Sovereign Identity Model for Agricultural Supply Chains

Abstract: The agricultural supply chain plays a critical role in ensuring food security and sustainability; however, it continues to face challenges related to data fragmentation, limited transparency, and insufficient trust among participating stakeholders. Existing supply chain systems are primarily based on centralized identity and data management models, which introduce single points of failure, restrict auditability, and raise privacy concerns. More critically, the absence of decentralized and stakeholder-controlled identity mechanisms limits accountability and verifiable governance across agricultural ecosystems. In this work, we present Ziraai SSI, a blockchain-based self-sovereign identity (SSI) prototype designed to support identity-centric governance and trust establishment in agricultural supply chains. The proposed approach integrates blockchain-based trust anchoring with self-sovereign identity principles using decentralized identifiers (DIDs) and verifiable credentials (VCs). Through this integration, stakeholders retain direct control over their digital identities while enabling cryptographically verifiable and privacy-preserving interactions across organizational boundaries. The system architecture follows a multi-layered design that addresses identity management, credential lifecycle handling, authentication, access control, and governance. To move beyond conceptual analysis, the framework is realized as a functional research prototype using standardized SSI technologies and an agent-based architecture. The implemented system supports end-to-end credential workflows, including secure connection establishment, credential issuance, selective disclosure, and proof-based verification, without reliance on centralized identity providers or authentication authorities. Experimental validation conducted in a controlled environment confirms correct execution of identity and credential lifecycles, decentralized authentication, and privacy-preserving verification. The evaluation focuses on functional validation within a controlled prototype environment and does not include large-scale scalability benchmarking. These results demonstrate the feasibility of identity-centric governance mechanisms in agricultural supply chains using standardized SSI technologies.

Author 1: Tariq Alar

Author 2: Gnana Bharathy

Author 3: Usha Batra

Author 4: Mukesh Prasad

Keywords: Blockchain; self-sovereign identity (ssi); agricultural supply chain; verifiable credentials; decentralized identity; privacy-preserving authentication

Paper 42: Rubric-Relational Discourse Modeling with Counterfactual Explainability for Multi-Trait Automated Essay Scoring

Abstract: Automated Essay Scoring (AES) systems often rely on holistic prediction and show weak alignment with rubric-based human evaluation. Existing deep learning approaches achieve moderate agreement but struggle to model discourse coherence and provide trait-faithful explanations. This study proposes a rubric-aware and discourse-faithful essay scoring framework that integrates contextual embeddings with sentence-level discourse modeling and rubric-specific attention. The framework generates both holistic and trait-level scores, while enabling counterfactual explanation of scoring decisions. Experiments conducted on the Learning Agency Lab – Automated Essay Scoring 2.0 dataset show that the proposed model achieves a Quadratic Weighted Kappa (QWK) of 0.86, Root Mean Square Error (RMSE) of 1.41, and Mean Absolute Error (MAE) of 1.12, outperforming CNN-LSTM, BERT-LSTM, and DeBERTa baselines. QWK evaluates ordinal agreement, while RMSE and MAE measure numerical prediction error. Trait-level performance reaches F1-scores of 0.89 for Content and 0.87 for Grammar, indicating strong rubric alignment. The proposed framework improves scoring reliability, interpretability, and consistency with human grading practices. It is suitable for large-scale educational assessment, formative feedback systems, and intelligent tutoring applications, offering a scalable and explainable solution for multi-trait essay evaluation.

Author 1: N. Sreedevi

Author 2: M. Madhusudhan Rao

Author 3: Sridevi Dasam

Author 4: Roopa Traisa

Author 5: Jasgurpreet Singh Chohan

Author 6: V. Saranya

Author 7: Ahmed I. Taloba

Keywords: Automated Essay Scoring; rubric-aware modeling; discourse representation; counterfactual explainability; multi-task learning

Paper 43: Face Sketch Recognition: Ethnic Groups Classification and Recognition Via a VGG16 Model Approach

Abstract: In the law enforcement investigation, the police use sketching techniques to identify suspects from an eyewitness's memory. Many automatic face sketch recognition systems that determine the perpetrator’s appearance from the face image datasets have been proposed. The aim is to conduct the arrest of the right offender. We propose this work to carry out a search based on the ethnicity criterion to speed up this automatic identification and to help authorities execute fast responses by launching the retrieval process only in a part of the dataset of face images. The goal of this study is to enhance the accuracy of ethnic face sketch classification by using the convolutional neural network built on the VGG16 architecture. The FairFace dataset, which includes seven ethnic face images: White, Black, Indian, East Asian, Southeast Asian, Middle Eastern, and Latino|Hispanic, was employed in the study. We convert the face images dataset to face sketch images, and we optimize the VGG16 model for seven classification outputs. This work shows that the VGG16 deep learning model offers a reliable, automated approach for ethnic face sketch classification and recognition. The used model achieved an accuracy reaching above 94% and produced a low false negative rate, which is crucial for minimizing undetected cases.

Author 1: Khalid OUNACHAD

Author 2: Mohamed EL GHMARY

Keywords: Forensic face sketch; ethnic classification; deep learning; transfer learning; optimized VGG16 model

Paper 44: Integrating Japanese Festival Traditions into Computational Thinking: A Culturally Responsive Approach to Micro:Bit-based Physical Computing

Abstract: The current pre-tertiary educational landscape in Japan is defined by a significant paradox. While technological infrastructure has reached unprecedented levels of saturation through the Global and Innovation Gateway for All (GIGA) School initiative, student engagement and qualitative learning outcomes in computational disciplines have shown signs of stagnation. This research addresses this disparity by proposing a framework grounded in Culturally Responsive Teaching (CRT) and Culturally Responsive Computing (CRC), utilizing the traditional Japanese festival game of Shateki (target shooting) as a primary pedagogical vehicle. By leveraging the BBC micro:bit and a modular physical computing architecture, this study demonstrates how situating abstract programming concepts, such as sequential logic, conditional branching, and signal modulation, within familiar “vernacular” and “heritage” cultural contexts can foster intrinsic motivation and improve academic achievement among K-12 students. Through a pilot evaluation conducted during a community festival, this study provides quantitative evidence that CRT-based interventions not only enhance technical proficiency but also bolster a sense of student agency and cultural belonging. The results suggest that the success of modern educational technology initiatives depends less on hardware distribution and more on the pedagogical translation of technical content into culturally meaningful experiences.

Author 1: Daiki Sugiyama

Author 2: M. Fahim Ferdous Khan

Author 3: Ken Sakamura

Keywords: Culturally responsive teaching (CRT); Culturally Responsive Computing (CRC); computational thinking; micro:bit; K-12 education; physical computing; STEM education

Paper 45: Machine Learning-Based Effort Prediction and Early Risk Detection in Software Development Projects: A Case Study

Abstract: Accurate effort estimation and early risk detection are critical for the success of software projects, as inaccurate forecasts can lead to schedule overruns, inefficient resource allocation, and unmet requirements. This study investigates the use of machine learning techniques to support task-level effort prediction and proactive risk identification in software project management. An applied case study was conducted on a simulated dataset of 500 software development tasks, described by planning, technical, and team-related features. Two ensemble-based regression models, Gradient Boosting and Random Forest, are evaluated for predicting actual task duration. Model performance is assessed using standard metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). To enable early risk detection, prediction errors are transformed into deviation-based indicators, and threshold-based classifiers are employed to identify tasks with moderate (>20%) and severe (>30%) schedule overruns. Confusion matrices and classification metrics are used to evaluate the effectiveness of the proposed alerting mechanism, and the distribution of high-risk tasks across sprint quantiles is analyzed to support managerial decision-making.

Author 1: Andreea-Elena Catana

Author 2: Adriana Florescu

Keywords: Software effort estimation; risk detection; machine learning; python; open data; gradient boosting; effort estimation; risk thresholding

Paper 46: Self-Supervised and Explainable Transformer-Based Architectures for Robust End-to-End Speech and Language Understanding

Abstract: The primary aim of this study is to meld self-supervised learning techniques with transparent transformer-based frameworks to enable resilient, end-to-end speech and language understanding, alongside pretraining deep transformer models using unannotated speech and text corpora. But the system's complicated structure makes it very hard to compute, and its ability to be understood depends in part on using rough benchmarks to judge feature relevance. This research work proposes an explainable, systematic transformer-based framework concept for understanding voice and language that integrates self-supervising learning with built-in explainability. The model proposed here presented a low word error rate, high accuracy, and interpretation on multiple datasets. The framework has many strengths, but it also has some challenges, which are highlighted in the work. This deep transformer architecture needs a lot of computing power, and figuring out how important something relies on indirect truth values. In the future, planned improvements include making the framework work with more than one language and more than one field, making transformer models work better in real time, and adding assessment methods that focus on human perspectives to make it even easier to understand. Subsequently, we will work on expanding into datasets that are multilingual and cross-domain, making more efficient forms of transformers for real-time use, and employing human-centered assessment to verify that we are interpreting things correctly in real time.

Author 1: Mahfuzul Huda

Keywords: Transformer models; self-supervised learning; explainable AI; speech recognition; natural language understanding; end-to-end systems

Paper 47: Secure Ultra-Low Latency Data Paths: A Hybrid Architecture for High Speed Networking under Adversarial Conditions

Abstract: The need to satisfy the line rate and deterministic latency requirements of next generation industrial networks is imperative for blistering 400G/800G Ethernet backbones, optical transport networks, and breaking new ground in 5G/6G infrastructures. Animals News: Other solutions limit attention towards detection accuracy or throughput separately without providing a discriminative systems-level architecture, leaving out bounded latency, scalable security as well as hardware-efficient adaptivity in an adversarial environment. We present a latency-constrained hardware-pipelined co-design framework, dubbed High Speed Secure Networking Architecture (HSSNA), capable of integrating probabilistic pre-filtering, adaptive lightweight cryptography, neural anomaly inference, and SDN-based routing into a single deterministic processing graph. In contrast to compositional security stacks, HSSNA defines the security-performance coupling as a constrained optimization problem that seeks to minimize the total processing delay without sacrificing the robustness of intrusion detection. Our contributions include (1) cross-layer security orchestration, which is embedded within the data path, (2) provable adversarial resilience guarantees, as a result of formally defined security properties, and (3) a parallel FPGA–GPU execution pipeline, which also removes sequential security bottlenecks. Through experiments conducted on a hybrid Mininet–NS3–FPGA/GPU testbed, we observe 60–77% lower latency, > 190 Gbps more throughput, and better robustness to real-time detection compared to conventional CPU-centric deployments. Our results prove HSSNA is systems-level re-architecture, high-speed secure networking, not a composition of prior art tools.

Author 1: Abdulbasid Banga

Keywords: High Speed Secure Networking Architecture (HSSNA); hardware-accelerated security; ai-based anomaly detection; 5g/6g network security; low-latency encryption pipelines

Paper 48: Hybrid Deep Learning for Academic Achievement Prediction Using Spatio-Temporal and Behavioral Data in Higher Education

Abstract: Accurate prediction of student academic performance is essential for enabling timely and effective educational interventions. Many existing prediction approaches focus either on academic outcomes or behavioral trends, without fully capturing the interaction between spatial performance indicators and their temporal evolution. To address this limitation, this study proposes a hybrid deep learning model that integrates spatio-temporal information for forecasting student achievement in higher education. The proposed framework combines a Convolutional Neural Network (CNN) to extract spatial features from normalized academic performance data with a Long Short-Term Memory (LSTM) network to model temporal patterns in student behavioral attributes, such as attendance and participation. In addition, FOX optimization is applied to adaptively tune the learning rate, improving training stability and predictive performance. The model is evaluated using student academic and behavioral datasets, and its performance is compared with commonly used baseline models. Experimental results show that the proposed CNN–LSTM approach achieves an accuracy of 97.18 per cent, outperforming standalone LSTM and Support Vector Machine (SVM) models. Furthermore, the model effectively classifies students into low, medium, and high academic risk categories, supporting early identification of at-risk students and facilitating timely intervention in higher education environments.

Author 1: Asim Seedahmed Ali Osman

Keywords: Hybrid deep learning; CNN; LSTM; student performance; FOX optimization

Paper 49: Farm and Learn: An Offline Mobile Learning System Integrating AR, AI, and Game-Based Learning for Agricultural Education Among Children

Abstract: This study presents Farm and Learn, an offline-first mobile learning system that integrates Augmented Reality (AR), Artificial Intelligence (AI), and Game-Based Learning (GBL) to enhance agricultural education among children in low-connectivity environments. Existing agricultural learning applications often provide isolated functionalities such as visualization or plant recognition, with limited pedagogical integration and insufficient support for rural deployment. To address these limitations, the proposed system combines immersive AR-based exploration, interactive gamified learning activities, and AI-assisted paddy plant growth-stage identification within a unified child-centered educational framework. The architecture adopts a modular offline-first design that enables core learning functionalities to operate without continuous internet access while allowing optional synchronization when connectivity is available. The AI component employs a lightweight YOLOv11n deep learning model validated through prototype inference to assess feasibility for future on-device deployment. The system was developed using Unity and ARCore and evaluated through user acceptance testing involving students, educators, and domain experts. Results demonstrate high usability, strong learner engagement, and improved learning performance, confirming the effectiveness of integrating immersive visualization, intelligent interaction, and gamified reinforcement in educational contexts. The findings highlight the practical potential of offline-first mobile learning platforms to support inclusive agricultural education and provide a scalable foundation for future intelligent educational systems in resource-constrained environments.

Author 1: Pubuditha De Silva

Author 2: Chamodya Prabodhani

Author 3: Daminda Herath

Keywords: Agricultural education; augmented reality; artificial intelligence; child-centered learning; game-based learning; offline mobile learning

Paper 50: An Enhanced Approach for Workmen’s Compensation Insurance Fraud Detection Based on Fuzzy Rule-Based System

Abstract: In Workmen’s Compensation insurance, fraud detection (FD) remains a significant challenge due to claims' inherent uncertainty and complexity. To address this, we propose an enhanced approach based on a fuzzy rule system (FRS) for FD. The FRS is designed to handle ambiguous and imprecise data, making it effective for identifying fraudulent patterns in insurance claims. Unlike traditional methods, the fuzzy system utilizes human-like reasoning by applying flexible rules to assess the likelihood of fraud under uncertain conditions. By modeling the decision-making process with fuzzy logic, the system allows for a detailed evaluation of claims, accommodating the gray areas that often exist in FD. This approach enables accurate and adaptive FD, reducing false positives and enhancing the precision of fraud identification. In imbalanced FD scenarios, the system achieves strong performance, such as an F1-score of 0.82 and MCC of 0.75, demonstrating its capability to correctly identify rare fraudulent cases despite class imbalance.

Author 1: Reham M. Essa

Keywords: Workmen’s compensation; fuzzy logic; fraud detection; rule-based system; insurance fraud; prediction

Paper 51: Design of a Vision Transformer-Based Architecture for Automatic Facial Emotion Monitoring in Workplace Environments

Abstract: Facial emotion recognition is increasingly considered in affective computing as a mechanism for unobtrusive emotional awareness in organizational environments. This study proposes the design of a Vision Transformer (ViT)-based system architecture for automatic facial emotion monitoring, focusing on deployability, modular integration, and data-governance considerations rather than model benchmarking. The architecture defines a complete pipeline comprising a visual data acquisition layer, a processing backend for transformer-based inference, and a web-based visualization interface intended for aggregated emotional analytics. Publicly available datasets such as FER2013 and AffectNet are identified as reference sources for model adaptation within the proposed framework. The work details system components, data flow, scalability strategies, and privacy-by-design mechanisms, including transient image handling and non-persistent processing. Rather than presenting experimental performance, this study provides a technical blueprint and feasibility analysis intended to guide future implementation and validation of transformer-driven emotion monitoring systems in workplace contexts. The proposed framework aims to bridge the gap between advances in deep learning models and their practical integration into real-world organizational infrastructures.

Author 1: Renzo Sebastian Gonzalez Caceres

Author 2: Jeramel Melissa Avila Saldaña

Author 3: Patricia Gissela Pereyra Salvador

Keywords: Facial emotion recognition; vision transformer; affective computing; system architecture design; workplace emotion monitoring; privacy-by-design; deep learning inference

Paper 52: Integrating Acoustic and Image Data Features for Melon Ripeness Classification Using Convolutional Neural Network

Abstract: This study evaluates three classification scenarios: image-based only, acoustic-based using Mel Frequency Cepstral Coefficients (MFCC), and a combined multimodal CNN architecture integrating both modalities. The experiments are conducted on a relatively small dataset comprising only 230 samples. To mitigate the risk of overfitting arising from the limited dataset size, data augmentation is applied to both image and audio data, with audio augmentation performed before the construction of the MFCC spectrogram. Experimental results demonstrate that the multimodal CNN with data augmentation achieves the best performance, with precision, recall, and F1-score, respectively, 0.95, 0.94, and 0.94. These results indicate that augmenting both image and audio data effectively enhances data diversity and model robustness, significantly improving classification performance. The findings confirm that combining complementary feature representations from multiple modalities with proper augmentation strategies substantially improves audio-visual classification tasks.

Author 1: Endang Purnama Giri

Author 2: Agus Buono

Author 3: Karlisa Priandana

Author 4: Dwi Guntoro

Keywords: CNN; multimodal learning; melon ripeness classification; image classification; acoustic classification; data augmentation

Paper 53: Comparative Analysis of Machine Learning Based Algorithms for Predicting Injury Severity in Road Accidents

Abstract: Road crash injury severity prediction is essential for intelligent transportation systems, yet challenged by severe class imbalance, rigid 4-class severity schemes (unhurt/slight/hospitalized/fatal), and optimal methodological selection. This study proposes a structured framework systematically evaluating four machine learning models—CatBoost, HistGradientBoosting, Random Forest, and SVM—across multiclass (native 4-class and ordinal wrapper), binary reduction (non-severe vs. severe), and oversampling techniques using crash data. Multiclass approaches reveal tree ensemble dominance but persistent rare severe class prediction difficulties. Binary class reduction substantially improves severe injury detection performance on this dataset by simplifying decision boundaries, while SMOTE oversampling provides algorithm-specific imbalance mitigation. Random Forest demonstrates the most stable binary performance across evaluation metrics, independent of oversampling strategies. This performance gain comes at the cost of reduced severity granularity compared to the original multiclass formulation. Overall, under imbalance-sensitive evaluation metrics, binary class reduction provides a pragmatic and operationally effective alternative to complex multiclass strategies for severe injury detection.

Author 1: Soumaya AMRI

Author 2: Mohammed AL ACHHAB

Author 3: Mohamed LAZAAR

Keywords: Machine learning; imbalanced data; road accident; multiclass classification; binary classification; injury severity prediction

Paper 54: Machine Learning and Deep Learning for Detecting Fake News in a Low-Resource Language

Abstract: Fake news detection has become a major problem in the digital age. This study presents an improved machine learning technique that achieves 91.99% accuracy in predicting fake news detection within Albanian textual datasets, demonstrating an improvement over existing baseline methodology. The implemented learning model uses 54 features that are specific to the Albanian language, such as red flags, credibility signals, punctuation patterns, and linguistic features. The model is tested on a balanced dataset of 3,994 news articles aggregated in Albanian from various sources. We compare it to several baselines, such as LSTM networks (80.35% accuracy) and BERT-augmented Naive Bayes classifiers (88.36% accuracy). In our Albanian dataset and experimental setting, XGBoost achieved 91.99% accuracy, indicating strong performance under the evaluated scenario.

Author 1: Elton Tata

Author 2: Jaumin Ajdarim

Author 3: Nuhi Besimi

Keywords: Fake news; Albanian language; machine learning; NLP

Paper 55: AMCS: Adaptive Multi-Controller SDN Security with Stateful Traffic Intelligence for Fast and Accurate Multi-Vector Attack Detection

Abstract: The rapid proliferation of Software-Defined Networking (SDN) in large-scale Internet of Things (IoT) ecosystems has amplified exposure to sophisticated, multi-vector cyberattacks that simultaneously exploit control- and data-plane asymmetries. Existing single-controller statistical detectors, while effective under high-volume anomalies, fail to sustain precision and responsiveness under dynamic, distributed, or low-rate attack conditions. Addressing this critical gap, we propose AMCS—an Adaptive Multi-Controller SDN Security framework that fuses stateful traffic intelligence with cooperative inter-controller decision-making to enable resilient, context-aware detection across complex IoT traffic. AMCS embeds lightweight, P4-based stateful processing directly in the data plane and augments it with adaptive entropy-driven anomaly evaluation and consensus-based coordination among distributed controllers. Extensive experiments on an SDN–IoT testbed demonstrate that AMCS achieves up to 99.7% detection accuracy for high-volume floods, 96.8% for low-rate anomalies, and 94.1% under mixed traffic, while maintaining false-positive rates below 5% and detection latencies as low as 1.24 s. The cooperative consensus protocol enhances cross-controller reliability to 98.4% with only 0.83 s synchronization delay, while reducing control overhead by 34.7% compared to the single-controller baseline. Moreover, the distributed mitigation layer reacts within 1.6 s on average, neutralizing over 97% of attack flows with negligible collateral impact. Collectively, these results confirm that integrating stateful in-switch analytics, adaptive thresholding, and multi-controller cooperation establishes a scalable, self-adaptive SDN security fabric—achieving both fast detection and stable defense against evolving multi-vector threats in IoT-driven networks.

Author 1: Ameer El-Sayed

Author 2: Mohamed Nosseir Hemdan

Author 3: Noha Abdelkarim

Author 4: Ehab Rushdy

Author 5: Hanaa M. Hamza

Keywords: SDN; IoT; multi-controller security; stateful traffic analysis; adaptive anomaly detection; distributed mitigation; entropy-based indicators; cooperative defense; P4 programmable networks; cyber-physical systems security

Paper 56: Wavelet-Based Dual-Domain Phase Alignment for Predicting Stock Index Trends from Investor Sentiment Cycles

Abstract: The dynamics in the financial markets are complicated and non-stationary, with a significant influence of the wave of investor sentiment. The recent changes in sentiment-based stock prediction have shown promising results, but the current research is to a large extent based on individual domain analysis, constant correlation, or the traditional machine learning framework, which constrains its capability to elucidate multi-scale temporal dynamics and phase-based lead-lag relationships. In order to overcome these weaknesses, a new Cross-wavelet Sentiment-driven Dual-domain Phase Alignment, abbreviated as CS-D²PA, is introduced for stock index trend prediction. The suggested structure has 91.8, 90.6, 92.1, 91.3, and 93.5 accuracy, precision, recall, F1-score, and trends consistency rate, respectively, proving to have a better predictive stability and a better classification performance in sentiment-driven stock trend forecasting. The non-stationary and multi-scale behavior of financial markets is dictated by non-periodic changes in investor sentiment cycles. This work proposes a Cross-wavelet Sentiment-based Dual-Domain Phase Alignment model (CS-D2PA) of predictive modeling of stock index trend. The framework combines the feature extraction by the continuous wavelet technique with the cross-wavelet phase difference estimation, as well as the structured alignment in time and frequency domains. The explicit modeling of the sentiment-price phase synchronization of the approach makes it possible to identify lead-lag interaction early and increases the predictability of the forecasts in volatile market conditions, which increases their interpretability.

Author 1: Monika Gorkhe

Author 2: Diksha Tripathi

Author 3: Pravin D Sawant

Author 4: Elangovan Muniyandy

Author 5: Veera Ankalu Vuyyuru

Author 6: Pratik Gite

Author 7: Raman Kumar

Keywords: Stock trend prediction; investor sentiment analysis; wavelet time–frequency analysis; phase synchronization modeling; financial market forecasting

Paper 57: Machine Learning-Based Autism Spectrum Disorder Classification Using an Enhanced Convolutional Neural Network Algorithm

Abstract: Autism Spectrum Disorder (ASD) is a complex neurological developmental disability that appears during early childhood. Conventional ASD diagnostic techniques rely on behavioural observations, characteristics, and clinical interviews. To overcome these limitations, numerous machine learning (ML) and Deep Learning (DL) techniques have been used to assist physicians. For the past three decades, biomedical images have been employed to diagnose neurodevelopmental disorders. The functional Magnetic Resonance (MR) images used in this study. This paper proposes a novel machine learning framework to classify ASD control from healthy controls. The proposed framework consists of two stages. In the first stage, an enhanced Convolutional Neural Network (CNN) is proposed to extract features. In the second stage, the extracted features are given to the machine learning classifiers. The proposed method is tested on the 1112 fMRI images. A total of 539 ASD participants and 573 healthy controls are included in this study. A total of 17 datasets from the ABIDE website are used. These datasets are collected from various international medical laboratories. The proposed framework outperforms the existing methods. The proposed algorithm achieved 92.45% across the entire ABIDE dataset and 98.61% on the individual dataset.

Author 1: P. Yugander

Author 2: M. Jagannath

Keywords: Autism; enhanced CNN; random forest; MR images; logistic regression

Paper 58: Privacy-Preserving Federated Learning for Multi-Institutional Lung Cancer Severity Detection

Abstract: Lung cancer continues to be the most common cause of cancer-related mortality globally and the timely detection of lung cancer and classification of its severity levels are critical to improving survival. Nonetheless, data privacy regulations and institutional data silos often create barriers to developing advanced robust AI models among clinical centers. This paper presents a framework for privacy-preserving multi-institutional lung cancer severity classification utilizing federated learning (FL) with secure aggregation, where only encrypted model updates are exchanged while raw patient data remain locally stored. The framework encompasses four new ideas: a privacy-preserving federated neural ensemble model (PP-FNE), a gradient boosting-(GB-)based FL strategy (MIF-GBF), a hybrid convolutional-transformer network (CF-CTN), and a semi-adaptive federated attention-aggregated model (SAFAM). Each of these ideas provides a way to connect sites in a multi-institutional effort while addressing data diversity/heterogeneity, model interpretability, and collaboration across sites while providing strong privacy protection measures for sensitive health data. The proposed framework is evaluated using a synthetic dataset developed to mimic the clinical heterogeneity of real-world clinical multi-site networks. The best-performing model, SAFAM, achieved an overall classification accuracy of 93.4%, demonstrated robustness to intelligently crafted noise (1.3% accuracy degradation), and preserved predictive performance under encrypted aggregation with minimal communication overhead per federation round. CF-CTN strengths lie in multimodal integration for lung cancer severity classification and model interpretability, while MIF-GBF had notable strengths in providing interpretability for GB-models specifically. PP-FNE exhibited stability as an ensemble model under variability across sites. All four individually based model FL methods restricted all communication/exchange across sites to conducting encrypted model update exchanges via a secure aggregation protocol, aligning with data minimization principles under HIPAA and GDPR). These results provide evidence that, if an FL approach considers task autonomous algorithmic innovations, accurate and privacy-protected lung cancer severity detection can be achieved through a distributed clinical setting.

Author 1: Ch. Srividya

Author 2: K. Ramasubramanian

Author 3: Myneni.Madhu Bala

Keywords: Federated learning; privacy-preserving AI; lung cancer detection; severity classification; multi-institutional data

Paper 59: RiskMIS: A Web-Based Risk Management Information System

Abstract: Risk management is the systematic process of identifying, assessing, monitoring, and responding to risks in projects and ongoing operations. Effective execution of risk management activities is essential for the successful completion of projects and the achievement of key performance indicators in operational environments. In recent years, organizations have increasingly emphasized the proactive identification and mitigation of risks before they materialize. Consequently, risk assessment is addressed early in the lifecycle and continuously revisited to ensure the proper functioning of ongoing operations and the successful delivery of projects. A central tool in the risk management process is the risk register, which consolidates all critical and relevant information related to identified risks. The risk register serves as a focal point around which risk management activities are organized. It is inherently dynamic and must be continuously updated to reflect changes in risk exposure and response strategies. When properly managed, the risk register supports informed decision-making and enables managers to handle risks in a systematic and timely manner. The primary objective of this study is to bridge theory and practice by developing a risk management information system centered on a comprehensive risk register. The study presents the design of the proposed system using established modeling techniques and diagrams, followed by the development of a functional prototype based on modern web technologies. Finally, a demonstration of the prototype is provided to illustrate its capabilities and practical applicability.

Author 1: Alya Thiab Almutairi

Author 2: Kassem Saleh

Keywords: Risk management; risk management information system; risk register; project risk; decision support systems

Paper 60: Uncertainty-Aware Volumetric Transformer with Dual Spatial-Channel Attention for Lung Nodule Classification

Abstract: Lung cancer is also among the most common causes of cancer-related deaths in the world, and the earliest possible detection of the cancer through computed tomography (CT) is important in the enhancement of patient survival. Nevertheless, accurate diagnosis is still a challenge as the nodules are small and indistinct, inter-rater consistency among radiologists, and the traditional deep learning systems have limited capacity to handle volumetric interactions and give interpretable and confidence-aware forecasts. This research suggests an uncertainty-cognizant Transformer-Enhanced Dual-Level Attention Network (TDA-Net) to classify lung nodules in CT images to deal with these issues. The suggested architecture combines a 3D Swin Transformer backbone and sequential spatial and channel attention fusion to be able to model both localized structural and global volumetric context. Moreover, Monte Carlo dropout is used in inference to measure predictive uncertainty and allows low-confidence cases to be identified and sent to a radiologist. The model is tested on a publicly available lung CT dataset, and it has an accuracy of 98.3% with high sensitivity to small nodules in the feature space. There is a separation of classes in the feature space, and the uncertainty rate is 5.1%. The findings of the experiment indicate that TDA-Net can be used as a supportive decision-making tool to diagnose lung cancer with the help of computers because it has better discriminative performance and uncertainty awareness when compared to the baseline models. Moreover, distinguishable uncertainty of predictions and uncertainty of models are present. Predictive uncertainty is measured through the variance of softmax probability distributions through stochastic forward passes, which is related to the ambiguity of data. Monte Carlo dropout is used to estimate model uncertainty as a Bayesian approximation, which represents parameter-level uncertainty due to a small amount of training data.

Author 1: B. N. Patil

Author 2: TK Rama Krishna Rao

Author 3: Nurilla Mahamatov

Author 4: Elangovan Muniyandy

Author 5: Arun Prasad.VK

Author 6: Chamandeep Kaur

Author 7: Aaquil Bunglowala

Author 8: Ahmed I. Taloba

Keywords: 3D Swin Transformer; dual-level attention mechanism; computed tomography; lung cancer diagnosis; uncertainty-aware deep learning

Paper 61: An Articulatory-Aware CNN-BiGRU-Attention Framework for Explainable Phoneme-Level Pronunciation Assessment in ESL Speech

Abstract: Proper pronunciation at the phoneme level has been known to be one of the most enduring problems affecting the Second Language learners of the English language (ESL) since the slight pronunciatory variations in the learned language may greatly influence its communicative power and the level of intelligibility. The existing methods of pronunciation evaluation, which are mostly made using automatic speech recognition (ASR), place their results at the word level or the sentence level and offer generic numerical scores with little linguistic meaning, which is not effective in assessing accented speech and subsequent correction. To overcome these shortcomings, the paper introduces an articulatory-conscious recognition model of phonemes that provides fine-grained and interpretable feedback to enhance ESL pronunciation. The novelty of the work is in the combination of a hybrid CNN-BiGRU-Attention architecture and an Articulatory Error Mapping Engine, which symbolically transforms phoneme-level articulation errors into articulatory errors, based on place of articulation, manner, voicing, and vowel quality articulatory deviations. The experimental analysis performed on the non-native English speech had a phoneme recognition accuracy of 91.4 that was much higher than the commercial ASR-based systems (78.3) and the traditional HMM-GMM baselines (70.5). The system was very sensitive to ESL pronunciation errors, making it 84 percent accurate in substitution, 82 percent accurate in deletion and 79 percent accurate in insertions in detection and articulatory mapping was over 87 percent accurate in all categories. The framework was tested in Python with deep learning packages and speech processing toolkits, and provided a scalable, explainable, and learner-focused system that can be used to support the intelligent training of ESL pronunciation and provide pedagogically significant feedback at the phoneme level.

Author 1: P. Bindhu

Author 2: Jasgurpreet Singh Chohan

Author 3: M. Durairaj

Author 4: Megha Sawangikar

Author 5: N. Neelima

Author 6: Elangovan Muniyandy

Author 7: G. Sanjiv Ra

Author 8: Loay F. Hussien

Keywords: Articulatory error analysis; attention mechanism; ESL pronunciation assessment; phoneme recognition model; speech processing framework

Paper 62: Cloud-Continuum-Based Deep Learning Optimization Framework for Next-Generation Healthcare Data Performance on IoT Platform

Abstract: The development of healthcare data performance analysis is becoming more driven by the incorporation of intelligent computing paradigms that guarantee real-time, scalable, and personalized feedback for coaches and athletes. However, existing healthcare data analytics systems are challenged with severe issues such as decision-making latency, processing capacity limitations at the edge, data fragmentation, and the inability to integrate across heterogeneous computing environments seamlessly. Athletic data in this scenario refers to a combination of biomechanical factors (motion capture, joint angles, gait patterns), biometric signals (heart rate, oxygen saturation, muscle activity), and sport-specific performance indicators (workload, speed, and acceleration). This paper introduces the Cloud-Continuum-based Deep Learning Optimization Framework (CC-DLOF). This novel architecture leverages the synergistic potential of edge, fog, and cloud computing to provide a dynamic and smart healthcare data performance on an IoT platform. CC-DLOF is a hierarchical continuum architecture, with real-time data gathering and lightweight analytics performed in the edge layer, contextual processing and federated learning in the fog layer, and global intelligence, deep model training, and long-term data storage in the cloud layer. A new Cloud-Fog-Edge Orchestration Device (CFEOD) provides dynamic allocation of computational tasks in terms of latency sensitivity and device capability. At the same time, a blockchain-supported access control is used to maintain data security and privacy. Simulation analysis, done in a simulated training environment that combines with real-world data sets, illustrates the performance of the framework in mitigating latency by 35%, increasing model accuracy by 22%, and boosting system scalability and reliability. CC-DLOF is a revolutionary way in healthcare data technology, leading to smart, responsive, and safe next-generation healthcare data performance on an IoT platform.

Author 1: G. Aravindh

Author 2: K. P. Sridhar

Keywords: Cloud; continuum; next-generation; healthcare data; performance; platforms; fog; edge; orchestration; device

Paper 63: Adaptive Network Security Framework for Distributed Quantum-Assisted Cloud Continuum Architectures

Abstract: Distributed Quantum-assisted Cloud Continuum architectures have brought revolutionary changes to the current computing environment through increased data proximity, reduced latency, and real-time responsiveness. These systems are architectures that integrate cloud, fog, and edge computing layers. However, this architectural development comes with many complicated security concerns such as heterogeneous devices, dynamic workloads, splintered attack surfaces, and different policy application across the spectrum. Such dynamic infrastructures need a new security paradigm that is resilient, scalable and responsive to guard against the aged and centralised safety measures. The Adaptive Threat-Aware Security Orchestration (ATASO) framework is a smart, context-aware, and scalable network security solution that is presented in this study as the way to overcome these challenges. Intelligent security monitoring layer (ISML) works in real-time and the Context-Aware Threat Analysis Engine (CTAE) detects distributed anomalies using federated deep learning. Adaptive Policy Enforcement Module (APEM) is a system based on context-aware and blockchain smart contracts to enforce mitigation policies. ATASO is made up of these three units. This multi-layer system is impregnable as far as enforcing policy enforcement is concerned and has a low latency overhead and the ability to monitor threats end to end, as dictated by its multi-layer design. The ATASO model is uniquely applicable where security responsiveness and low-latency response is of utmost importance, including health care monitoring networks, autonomous vehicle networks, smart city networks, and industrial IoT networks. By conducting extensive simulation studies, the approach has been discovered to outperform the existing approaches in a number of important dimensions including the detection accuracy (more than 96%), the response latency (up to 40% less), and the resource consumption where large computers are involved. These findings confirm that ATASO has the potential of being a sophisticated adaptive security system that will protect future cloud continuity designs against the developing cyber threats.

Author 1: P. Suseendhar

Author 2: K. P. Sridhar

Keywords: Adaptive; network; security; distributed; cloud; continuum; threat; aware; orchestration; intelligent; engine; policy; enforcement module

Paper 64: Deep Hybrid Learning for Sustainable Industrial Forecasting: Integrating CNN–LSTM Models to Enhance Economic Efficiency and Carbon Performance

Abstract: This paper explores the contribution of neural network-based safeguarding models to enhancing the environmental resilience and economic efficiency of industrial supply chains. The methodology includes a review of existing literature for a quasi-experimental study conducted from the perspective of a manufacturer. Using this approach, the study analyzes the transition from traditional statistical safeguarding practices to modern neural predictive frameworks, the amount of data available, and assesses their impact on decision-making and overall chain performance. The results from a Tunisian organization indicate that deep hybrid training architectures, particularly CNN-LSTM models, significantly improve the accuracy of demand forecasting, resulting in concurrent gains in operational efficiency and environmental performance. The organization also achieved a reduction in its annual costs of 2.25 million Tunisian dinars, leading to a decrease in carbon emissions. The study also identifies key obstacles, such as the fragmentation of data infrastructure, the lack of digital skills, and global development costs, which necessitate the effective adoption of deep training. Based on these findings, the paper proposes a dual-performance neural network framework to help managers and policymakers align technological innovation with the realities of emerging economies.

Author 1: Mohamed Amine Frikha

Author 2: Mariem Mrad

Author 3: Younes Boujelben

Author 4: Soufiene Ben Othman

Keywords: Neural networks; deep learning; CNN-LSTM; AI-enabled demand forecasting; sustainable supply chain; economic performance; digital transformation; data quality; emerging economies

Paper 65: A System-Oriented Machine Learning Approach for Planning and Execution of Decisions in Software Project Management

Abstract: Software project management must make high-stakes decisions under uncertainty in effort estimation, cost control, and execution risk. Although machine learning has enhanced predictive accuracy, several studies employ it only in isolated planning or execution tasks, thereby limiting its usefulness as an end-to-end decision support system. This study describes DeciBoost PM, a single framework that facilitates both planning-layer estimation and execution-layer delay-risk management using a single CatBoost backbone with inherent interpretability. On three heterogeneous public datasets, namely Desharnais to estimate effort, PROMISE to estimate cost, and an Apache JIRA issue-tracking dataset to classify delays and risks, we assess the framework. The same pipeline is used on the datasets, such as preprocessing, feature engineering, leakage-conscious splitting, and equal validation. Standard regression and classification measures are used to measure performance and compare the results with baseline learners. Findings indicate that DeciBoost PM has good and consistent predictive performance with low variance among tasks, thus enhancing better estimation accuracy and delay-risk discrimination. The framework provides transparency in its explanations of SHAP-based and threshold-controlled decision rules that can be directly translated to actionable managerial indicators. In general, DeciBoost PM has captured machine learning as a system-level, practical decision support methodology throughout the software project life cycle.

Author 1: Foziah Gazzawe

Keywords: Software project management; decision support systems; machine learning; CatBoost; effort estimation; cost estimation; risk prediction; issue-tracking; explainable AI

Paper 66: Cluster Domain-Aware Client Selection for Federated Learning in The Healthcare Field (CDCSF)

Abstract: Client selection remains a critical challenge in Federated Learning (FL). Resource-aware strategies aim to reduce training delays and mitigate stragglers by selecting an appropriate subset of clients in each round. However, these methods prioritize computationally strong clients and exclude resource-constrained clients. In healthcare settings, this approach is impractical because it removes entire domains from training, which harms generalisation. To address these challenges, we propose CDCSF, a domain-aware client selection framework that re-partitions clients into domain-homogeneous groups in each iteration. CDCSF is a dynamic clustering framework based on the (EM) algorithm that clusters clients based on local feature prototypes to enhance domain diversity. The framework incorporates a reliability score derived from an exponential moving average of training time to favor efficient clients. Simultaneously, a fairness score is introduced to ensure that underrepresented clients can still contribute to the training. This approach preserves sufficient representation across all domains to improve model generalization and accelerate convergence. We conduct extensive experiments on a healthcare benchmark dataset to validate the effectiveness of CDCSF. The proposed method improves accuracy by 2% over FedAvg under domain shift and outperforms PoC by 8%. With the proposed adaptive client selection strategy, we further demonstrate that CDCSF converges significantly faster than baseline methods under heterogeneous resource and data conditions.

Author 1: Sanaa Lakrouni

Author 2: Marouane Sebgui

Author 3: Slimane Bah

Keywords: Federated learning; healthcare; distributed learning; data heterogeneity

Paper 67: Temporal Attention Networks for Real-Time Multimodal Emotion Recognition from EEG and fNIRS Signals

Abstract: Emotion recognition is critical in the development of real-time mental health care and individualized cognitive behavior. Current strategies to recognize cognitive emotions frequently fail to capture complex time dependencies and multimodal physiological reactions, leading to sub-optimal performance and inaccurate generalization. To overcome such shortcomings, the proposed study suggests TCADNet, a new deep learning model that integrates Temporal Convolutional Networks (TCN), attention-based feature weighting, and GAN-based data augmentation to achieve a high recognition rate of the emotional states through EEG and fNIRS recordings. The model utilizes the TCNs to extract both short-term and long-term temporal trends, and the attention mechanism emphasizes salient parts that bring about emotions, which improves interpretability. Moreover, a Deep Convolutional GAN creates artificial signals of unrepresented emotion classes, eliminating data imbalance and enhancing generalization. The TCADNet model is coded in Python on the TensorFlow/Keras system, and its key components are preprocessing, time modeling, attention weighting, data augmentation, and last classification by SoftMax layers. Experimental outcomes indicate that TCADNet has high recognition performance, with overall recognition, accuracy, precision, and recall, and F1-scores of over 98, which is higher than conventional CNN, LSTM, and separate TCN models. The suggested methodology can be useful to researchers, clinicians, and mental health professionals as it allows them to monitor cognitive and emotional conditions in real-time with a reliable, decipherable, and scalable instrument and provides an opportunity to detect and respond to the issue promptly and implement a tailored intervention plan in educational or health-related settings.

Author 1: Vinod Waiker

Author 2: Anne Marie D. Pahiwon

Author 3: Ankush Mehta

Author 4: Gadug Sudhamsu

Author 5: Pavithra M

Author 6: K. Kiran Kumar

Author 7: B Kiran Bala

Author 8: Osama R. Shahin

Keywords: Emotion recognition; Temporal Convolutional Network; attention mechanism; mental healthcare; EEG

Paper 68: Explainable AI with DRL for Smart Home Energy Management and Residential Cost Savings in IoT-Based Autonomic Systems

Abstract: As more people seek ways to improve their homes and workplaces while reducing energy consumption, smart home systems are becoming increasingly prevalent. Unfortunately, the complexity and "black-box" nature of these systems made it difficult to deploy AI-enabled decision-making simulations, raising issues with explainability, confidence, transparency, responsiveness, and fairness. Explainable Artificial Intelligence (XAI), a rapidly developing discipline, addresses these problems by offering justifications for various decisions and behaviors of the systems. This research describes a novel method for IoT-based autonomic devices to control energy that combines XAI with Deep Reinforcement Learning (DRL) to achieve significant Home Energy Management System (HEMS) for household cost reductions. The proposed approach leverages XAI's features to improve the accessibility and transparency of DRL agents, helping consumers understand and trust autonomous power management decisions. By optimizing energy usage patterns and adapting to changing environmental conditions, the proposed solution ensures effective energy use while maintaining user comfort. Use in-depth modeling and real-world applications to demonstrate the solution's efficacy, highlighting its potential to reduce energy consumption costs and promote sustainable living. This study sets a new standard for clarity and flexibility in AI-driven smart home systems, paving the way for more reliable and user-friendly IoT software. It is important to note that developing a thermal dynamics model and understanding unidentified variables are not prerequisites for the proposed technique. Results from simulations based on real-world data show the resilience and effectiveness of the recommended strategy.

Author 1: Prabagaran A

Author 2: Srinivasan Sriramulu

Author 3: S. Premkumar

Author 4: Sheelesh Kumar Sharma

Author 5: Karthi Govindharaju

Author 6: Angelin Blessy J

Keywords: Explainable artificial intelligence; smart home energy management; deep reinforcement learning; Internet of Things; energy optimization

Paper 69: Securing Blue Carbon Accounting: A Cryptographic Framework for Coastal Ecosystem Monitoring

Abstract: Blue carbon ecosystems have significant long-term carbon sequestration capacity, making them an important nature-based solution for climate change mitigation. However, monitoring and accounting processes are increasingly dependent on distributed IoT sensors, satellite remote sensing, and cloud-based analytics platforms. This growing digitalization exposes the blue carbon data lifecycle to risks such as tampering, unauthorized access, and loss of data provenance. A compliance-aware cryptographic framework is presented to secure blue carbon accounting throughout the end-to-end process, from in situ measurement to carbon credit verification. In contrast with generic IoT–blockchain architectures, the framework binds sensing devices to national Public Key Infrastructure (PKI) identities and produces audit-ready cryptographic evidence aligned with Monitoring, Reporting, and Verification (MRV) workflows. The design employs the Elliptic Curve Digital Signature Algorithm (ECDSA) to ensure authenticity and non-repudiation, Advanced Encryption Standard in Galois/Counter Mode (AES-GCM) encryption for confidentiality, a hash-chained log for ordered integrity, and Secure Multiparty Computation (SMC) for privacy-preserving validation. Experimental results under simulated attacks (n = 50) demonstrate a 100% detection rate across the evaluated tampering scenarios, while maintaining an average IoT-layer cryptographic latency below 10 ms and a blockchain throughput of 145 transactions per second, exceeding the requirements for continuous ecosystem monitoring. These findings indicate that strong lifecycle-wide cryptographic guarantees can be achieved without imposing prohibitive computational overhead.

Author 1: Heider A. M. Wahsheh

Keywords: Blue carbon accounting; environmental cybersecurity; blockchain security; Public Key Infrastructure (PKI); elliptic curve cryptography; secure multiparty computation; IoT security; data integrity; carbon credit verification; MRV

Paper 70: Privacy-Aware Customer Segmentation Using a Distributed Graph-Based Attribute Projection Framework

Abstract: Customer segmentation plays a vital role in Business Intelligence (BI) by enabling organizations to understand customer behavior, enhance personalization, and support informed decision-making. Conventional segmentation approaches, including K-Means clustering, hierarchical methods, and hybrid deep learning models, often face limitations when handling high-dimensional customer data and typically lack built-in mechanisms to address privacy concerns. As customer analytics increasingly relies on sensitive personal information, these limitations pose significant challenges for responsible data-driven applications. To overcome these issues, this study introduces a Distributed Graph-Based Attribute Projection Framework (GAPF) for privacy-aware customer segmentation. The key novelty of the proposed framework lies in its ability to minimize sensitive attribute exposure while preserving meaningful relational patterns among customers through graph-based representations. GAPF employs a distributed processing pipeline that integrates attribute projection to reduce identifiability, heuristic-driven customer similarity graph construction, graph convolutional network (GCN)–based feature learning, and community detection for final segmentation. The framework is implemented using Python, NetworkX, and PyTorch Geometric and evaluated on the Mall Customers dataset and large-scale anonymized synthetic data to assess scalability. Experimental results demonstrate that GAPF achieves superior segmentation performance, with an accuracy of 98%, precision of 92.5%, recall of 94.0%, and an F1-score of 93.2%, while also exhibiting efficient execution and reduced privacy risk. These findings confirm GAPF as a robust and practical solution for privacy-aware BI applications.

Author 1: Pentareddy Ashalatha

Author 2: G. Krishna Mohan

Keywords: Business intelligence; privacy-aware customer segmentation; graph-based learning; federated analytics; graph neural networks

Paper 71: IoT-Driven Sensor Selection for Smart Water and Electricity Systems: A TOPSIS and VIKOR Approach

Abstract: In the era of smart cities and intelligent resource management, the need for precise and efficient sensing devices is increasingly critical. Among essential infrastructures are water and electricity systems, which necessitate continuous monitoring to optimize consumption, detect anomalies, and reduce operational costs. Several commercial sensors for water metering and electricity metering are available on the market, each differing in precision, power consumption, communication protocols, and durability. This diversity creates a decisional challenge when choosing the best sensor for a certain application. We report herein a comparative study of various water and electricity sensors by means of a multi-criteria decision-making approach. We adopt TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) and VIKOR (VIseKriterijumska Optimizacija I Kompromisno Resenje), two robust quantitative MCDM techniques. The carefully selected set of evaluation criteria and a wide range of sensors commonly used were analyzed. It turned out that the ranking was different based on context and application; therefore, valuable insights have been provided for engineers, system designers, and decision-makers in the IoT domain. It is expected that the present study will be of practical help for choosing the best sensors in smart metering applications and prove the efficiency of integrated decision models in technical evaluations.

Author 1: Oumaima Rhallab

Author 2: Rachid Dehbi

Author 3: Zouhair Ibn Batouta

Author 4: Amine Dehbi

Keywords: Water and electricity system; water metering; electricity metering; MCDM techniques; TOPSIS; VIKOR; IoT; smart metering

Paper 72: A Feasibility Study on Synthetic RGB-NIR Image Generation for Oil Palm Fresh Fruit Bunch Grading

Abstract: Accurate ripeness of grading oil palm fruit bunches (FFBs) is essential for optimizing oil quality and harvesting decisions. While near-infrared (NIR) imaging provides useful spectral cues for ripeness assessment, its adoption in field conditions is limited by sensor cost and system complexity. This study presents a low-cost alternative by generating synthetic NIR images from RGB inputs using a U-Net-based image translation model and integrating the generated NIR with RGB channels for ripeness classification. Five deep learning models, including a custom CNN, ResNet-50, EfficientNet-B0, DenseNet-201 and MobileNetV3, were evaluated under RGB-only and RGB + synthetic NIR configurations using identical training protocols. Experimental results demonstrate consistent performance improvements when synthetic NIR was incorporated. EfficientNet-B0 achieved the highest overall accuracy of 90.3%, while MobileNetV3 obtained the highest macro-averaged F1-score of 85.4%, indicating strong and balanced classification across ripeness classes. Confusion matrix analysis further revealed complementary strengths between the models, where EfficientNet-B0 showed stronger robustness in late-stage maturity detection, and MobileNetV3 provided improved discrimination of early-stage ripeness. The results demonstrate that synthetic NIR augmentation enhances classification performance and training stability without requiring specialized imaging hardware.

Author 1: Nor Surayahani Suriani

Author 2: Norzali Hj Mohd

Author 3: Shaharil Mohd Shah

Author 4: Siti Zarina Muji

Author 5: Fadilla Atyka Nor Rashid

Keywords: Generative AI; deep learning; U-Net image translation; EfficientNet-B0; MobileNetV3

Paper 73: The Acquiring Optimal Models of Random Forest and Support Vector Machine Through Tuning Hyperparameters in Classifying the Imbalanced Data

Abstract: Machine learning models most often misclassify the positive class in the dataset with class imbalance. Besides, a sophisticated model involves the hyperparameters that need to be tuned to the optimal values. The study aims to tune hyperparameters of random forest (RF) and support vector machine (SVM) models using 5-fold cross-validation data, to build the best RF and SVM for two data scenarios: the original and oversampling training data, and to compare the models' performances in either the training or testing data. The RF hyperparameters: the instance number in the leaf node and tree depth of the RF, were acquired (500, 10), respectively. Whereas, the SVM hyperparameters: the values of gamma and constant, were acquired (0.001, 500), respectively. The benchmark models achieved around 98% across the accuracy, precision, recall, and F1 score metrics. However, it performed worse on the Mathew's Correlation Coefficient (MCC) and Area Under the Curve (AUC): 0.0000 and 0.5000, respectively. The models trained on the class-imbalance dataset failed to predict the positive class. Although the best RF and SVM models trained on the oversampled dataset perform worse than both benchmark models across four standard metrics, the RF best model shows improvements of approximately 7% (from 0.000 to 0.067) and 11% (from 0.500 to 0.612) while the SVM best model show slightly different improvements of approximately 6% (from 0.000 to 0.056) and 11% (from 0.500 to 0.611) in MCC and AUC, respectively. Both the RF and SVM models improve in predicting the positive class, and the best RF model performs slightly better.

Author 1: Dwija Wisnu Brata

Author 2: Arif Djunaidy

Author 3: Daniel Oranova Siahaan

Author 4: Samingun Handoyo

Keywords: Area under the curve; cross-validation folds; Matthew's correlation coefficient; optimal hyperparameters; oversampling technique

Paper 74: K-Nearest Neighbors Algorithm for Short-to-Medium Term Directional Stock Price Forecasting: An Analysis of Thailand’s Banking Sector

Abstract: This research explores the efficacy of a parsimonious K-Nearest Neighbors (KNN) framework for short-to-medium term stock direction forecasting, focusing specifically on the banking sector within Thailand’s SET50 index. Prior preliminary analysis, aimed at determining the optimal prediction horizon, indicated that a 60-day forecast yielded the most effective results, establishing the scope of this study as medium-term prediction. The objective of this analysis is to determine if a 60-day directional movement can be effectively captured using a minimalist feature set limited to the current day’s Opening Price and the previous day’s 14-day Simple and Exponential Moving Averages. Employing a rolling-window validation methodology on seven key banking stocks during H1 2025, the KNN model demon-strated significant predictive capability. The average accuracy across the selected banking stocks reached 82.0%, with standout performance for TISCO and SCB. While results varied across stocks, our findings substantiate the theoretical and practical sufficiency of a simplicity-first approach. The research demonstrates that in high-noise emerging markets, feature sparsity and instance-based logic serve as an essential defense against overfitting, providing institutional practitioners with a transparent and robust alternative to complex methodologies.

Author 1: Passawan Noppakaew

Author 2: Parit Wanitchatchawan

Author 3: Kanchana Phuhoy

Author 4: Natthasorn Seubwong

Keywords: Stock trend prediction; K-Nearest Neighbors; technical analysis

Paper 75: Multi-Criteria Methodology for Selecting Communication Protocols in M2M Environments

Abstract: The continuous expansion of Machine-to-Machine (M2M) communication and Internet of Things (IoT) ecosystems has significantly increased the complexity of selecting appropriate communication protocols and data flow management systems. Contemporary M2M deployments operate across heterogeneous functional domains, including sensor networks, transactional systems, and real-time streaming environments, each imposing distinct and often conflicting non-functional requirements such as latency, reliability, scalability, and resource efficiency. This study proposes a domain-oriented multi-criteria decision-making methodology for structured protocol selection in M2M environments. The framework integrates the Analytic Hierarchy Process (AHP) for context-dependent weighting of evaluation criteria with the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) for quantitative ranking of alternative technologies. A structured domain taxonomy is introduced to dynamically align evaluation priorities with functional deployment characteristics, and 99th percentile latency (Lp99) is incorporated as a primary performance indicator to capture tail-behavior effects critical for M2M reliability. Beyond ranking computation, the methodology formalizes a reproducible analytical workflow linking empirical measurements to domain-specific decision out-comes and incorporates a sensitivity-analysis perspective to assess ranking robustness under variations in criterion weights. The proposed framework establishes a transparent and adaptable decision-theoretic foundation for context-aware communication protocol selection in heterogeneous M2M scenarios.

Author 1: Oleg Iliev

Keywords: Machine-to-Machine communication; Internet of Things; multi-criteria decision making; AHP; TOPSIS; communication protocol selection; non-functional requirements

Paper 76: Generative AI as a Catalyst for Interoperability and Data-Driven Decision Support in Healthcare Systems of Developing Countries

Abstract: Interoperability across heterogeneous information systems remains a persistent challenge, particularly in resource constrained contexts where infrastructures are fragmented and data formats remain incompatible. This study introduces a novel methodology that integrates generative Artificial Intelligence (AI) with the urbanization of information systems to enable scalable and seamless interoperability. The approach employs AutoGen AI, an open-source orchestration framework powered by Large Language Models (LLMs), specifically GPT-4o, to coordinate task-specific intelligent agents for data extraction, transformation, and harmonization. By converting disparate data into a standardized JSON representation, the architecture resolves both syntactic and semantic inconsistencies while simultaneously emphasizing multi-agent concurrency, distributed orchestration, and computational scalability, resulting in improved through-put, reduced latency, and enhanced robustness. A real-world healthcare case study is presented to illustrate the framework’s effectiveness: heterogeneous clinical datasets were unified into a coherent JSON structure, enabling accurate health indicator generation and reliable decision support. Experimental results demonstrate substantial improvements in system connectivity, processing efficiency, and integration reliability, with potential to generalize far beyond the medical sector. Moreover, the methodology incorporates advanced prompt engineering and context-aware dialogue design, minimizing model hallucinations and ensuring trustworthy outputs in LLM-driven processes. Overall, the study positions generative AI not only as a promising solution for interoperability in health informatics, but also as a transformative paradigm for intelligent system integration across diverse domains characterized by distributed, heterogeneous environments.

Author 1: YNSUFU Ali

Author 2: MOSKOLAI NGOSSAHA Justin

Author 3: AYISSI ETEME Adolphe

Author 4: BOWONG TSAKOU Samuel

Keywords: Decision support systems; interoperability; generative Artificial Intelligence; Large Language Model (LLM); heterogeneous data sources; information system

Paper 77: Development of Lightweight Residual Convolutional Neural Network for Efficient Facial Emotion Recognition

Abstract: Facial Emotion Recognition (FER) is essential for successful human-computer interaction; however, deploying robust systems on edge devices remains difficult. Recent techniques, such as Vision Transformers (ViTs) and deep ensemble networks, have achieved high accuracy, but suffer from extreme computational overhead and high latency, making them unsuitable for real-time use on limited hardware. The primary challenge lies in maintaining high discriminative power while operating under strict memory and power constraints. To address this, the objective of this research is to develop an efficient Residual Convolutional Neural Network (CNN) optimized for CPU-based inference. The proposed architecture utilizes a hierarchical structure, integrating three consecutive residual blocks with progressively increasing filter depths of 32, 64, and 128. These are engineered to enhance gradient flow and refine feature representation from low-resolution (48 × 48) grayscale images. Comprising only 552,455 parameters and achieving a 12.4 ms latency on standard CPUs, the model balances efficiency and performance. Experimental results on the FER2013 dataset reveal a classification accuracy of approximately 71.4%, outperforming several existing lightweight frameworks. A comprehensive assessment using confusion matrices and ROC curves validates the architecture as a practical solution for real-time affective computing on resource-constrained devices.

Author 1: Yelnur Mutaliyev

Author 2: Zhuldyz Kalpeyeva

Keywords: Facial Emotion Recognition; residual neural networks; lightweight convolutional neural networks; affective computing; facial expression recognition 2013 dataset; CPU-optimized architecture; pattern recognition

Paper 78: Integrating Deep Reinforcement Learning for Initialization and Adaptive Pheromone Updates in Ant Colony Optimization for UAV Pathing

Abstract: Unmanned Aerial Vehicles (UAVs) are indispensable assets for missions in dynamic and complex environments, requiring highly efficient path planning that simultaneously optimizes the often-conflicting objectives of minimizing flight distance, energy consumption, and mission time. While Ant Colony Optimization (ACO) is a recognized and effective metaheuristic for this domain, its performance is significantly constrained by a static, empirically-derived pheromone update mechanism, which prevents the algorithm from adaptively learning or optimally managing the search process. To overcome this critical limitation, this study introduces a novel DRL-Assisted ACO framework where a Deep Reinforcement Learning (DRL) agent is seamlessly integrated with the ACO to strategically determine the optimal paths under multi-objective constraints. This intelligent agent is tasked with learning the optimal, mission-specific pheromone update strategy. It achieves this by observing the performance of generated paths and receiving a sophisticated reward signal meticulously derived from the Analytic Hierarchy Process (AHP), which systematically weights the mission objectives. Validated through a simulated case study conducted in Khartoum State, Su-dan, the DRL-Assisted ACO approach has demonstrably achieved superior performance, exhibiting marked gains in convergence speed and generating paths with a significantly higher overall multi-objective utility score, thereby delivering a robust and adaptive solution essential for high-stakes autonomous UAV operations.

Author 1: Mohamed A. Damos

Author 2: Wenbo Xu

Author 3: Abdolraheem Khader

Author 4: Ali Ahmed

Author 5: Mohammed Al-Mahbashi

Author 6: Almuhannad S.Alorfi

Keywords: Deep Reinforcement Learning; Ant Colony Optimization; adaptive pheromone update; UAV pathing

Paper 79: A Context-Aware Hybrid Recommendation Framework for E-Learning Platforms

Abstract: E-learning platforms provide learners with ex-tensive digital resources that enable self-paced and location-independent study. However, the overwhelming volume of learning materials offered by a wide range of institutions and content providers makes personalized guidance increasingly essential for effective knowledge acquisition. As a result, recommender systems have become fundamental components of modern e-learning environments, helping to reduce information overload and support individualized learning experiences. In general, the richer and more diverse the available data, the more accurate and relevant the resulting recommendations. Despite these advantages, conventional recommendation approaches often fail to fully exploit the contextual and relational information inherent in e-learning ecosystems, which limits their adaptability and predictive precision. This study proposes a hybrid recommendation framework that integrates collaborative filtering, content-based filtering, and context-aware modeling to generate more accurate and adaptive course recommendations. The proposed system infers learner preferences by combining historical interaction data, contextual attributes, and course characteristics, while also incorporating temporal and environmental factors that influence learning behavior. Experimental evaluations based on SVD, TF-IDF, and RNN models applied to a well-established benchmark dataset demonstrate that the proposed hybrid framework significantly improves recommendation accuracy, coverage, and adaptability compared with baseline methods. Furthermore, the integration of contextual information effectively alleviates the cold-start problem and better captures learners’ evolving goals and learning trajectories. Overall, the results confirm that combining multiple recommendation paradigms within e-learning platforms enables more adaptive, personalized, and scalable learning pathways, making the proposed system suitable for diverse educational contexts and learner profiles.

Author 1: Kaoutar Errakha

Author 2: Amina Samih

Author 3: Sanaa Dfouf

Author 4: Abderrahim Marzouk

Keywords: E-learning; recommender systems; hybrid approach; personalized learning; course

Paper 80: Efficient CNN-Based Time-Domain Denoising of Impulsive Noise in NB-PLC Systems

Abstract: In this study, a convolutional neural network (CNN)-based time-domain denoising approach is proposed to suppress impulsive noise which is considered as the most sever impairments in narrowband powerline communications (NB-PLC). Unlike conventional techniques, such as clipping and blanking, the proposed method does not require prior knowledge of noise statistics. The introduced CNN network is trained using synthetically generated OFDM signals corrupted by Middleton Class-A impulsive noise, calibrated from real NB-PLC measurement data. Extensive G3-PLC-compliant simulations demonstrate that the proposed method significantly outperforms classical blanking and clipping schemes. At an SNR of 10 dB, the proposed CNN achieves a mean squared error (MSE) of 1.2×10−4, compared to 2.3×10−4 and 2.5 × 10−4 for blanking and clipping, respectively, under time-varying impulsive noise conditions. Moreover, the receiver incorporating the denoising method closely approaches the ideal AWGN reference under low impulsive noise density and for SNR values above 12 dB.

Author 1: Wided Belhaj Sghaier

Author 2: Fatma Rouissi

Author 3: Héla Gassara

Author 4: Fethi Tlili

Keywords: NB-PLC; Middleton Class-A; OFDM; impulsive noise; deep learning; CNN

Paper 81: Leveraging Attention Mechanism and Class Weighting for Legal Event Detection in Chinese Text

Abstract: Event detection is an information extraction task that involves extracting specified event types from textual sequences. Currently, most event detection studies focus on English corpora; there is a lack of exploration in other linguistic contexts. Thus, a study on event detection in the Chinese corpus is essential. Sequence-based event detection has been extensively studied in the past, and many studies have utilized high-performance neural network models, such as traditional recurrent neural networks. This study aims to enhance the performance of sequence models by altering the base model, Bidirectional Long Short Term Memory (BiLSTM), to a Bidirectional Gated Recurrent Unit (BiGRU) and incorporating multi-head attention mechanisms, Conditional Random Fields (CRF), and class weights. These modifications not only improve the model’s accuracy but also enhance computational efficiency by reducing the number of parameters relative to large pre-trained models such as Bidirectional Encoder Representations from Transformers (BERT). The experimental findings demonstrate that the proposed model’s modification achieves an F1 Score of 83.55 for the micro standard and 78.07 for the macro standard. This presents a substantial improvement over the baseline, delivering performance nearly on par with state-of-the-art BERT-based models on the same dataset, while requiring significantly fewer parameters.

Author 1: Jinhong Hu

Author 2: Shaidah Jusoh

Keywords: Event detection; information extraction; Chinese corpus; recurrent neural network

Paper 82: SIFChain: A Decentralized Framework for Secure Storage Sharing and Dynamic Access Control in Virtual Power Plants

Abstract: Virtual Power Plants (VPPs) face significant challenges in secure data management and sharing, including risks of centralized control, single points of failure, and dynamic access requirements among multiple stakeholders. To address these issues, this study proposes SIFChain, a decentralized framework that integrates Hyperledger Fabric, the InterPlanetary File System (IPFS), and a revocable Ciphertext-Policy Attribute-Based Encryption (CP-ABE) scheme with collaborative key generation. Unlike existing solutions such as Filecoin or Storj, SIFChain introduces a dual-channel blockchain architecture that separates public operational data from sensitive attribute information, mitigating privacy leakage and access policy exposure. The framework achieves fine-grained, dynamic access control with forward and backward security through an enhanced CP-ABE mechanism. Experimental evaluation demonstrates that SIFChain provides scalable performance: data upload/download times in-crease linearly from 1 MB to 1 GB, blockchain transaction latency remains under 5 ms for typical operations (registration, access requests, policy updates), and attribute-based encryption/decryption overhead scales linearly with policy complexity. These results confirm the practicality of SIFChain for secure, cross-organizational data sharing in Virtual Power Plant ecoSystems.

Author 1: Xiaochuan Xu

Author 2: Xiao Xin

Author 3: Jie Liu

Author 4: Ruiqi Fang

Author 5: Dekai Liu

Author 6: Zhixin Li

Keywords: Virtual Power Plant; access control; CP-ABE; IPFS; blockchain; distributed energy; privacy protection

Paper 83: AI-Driven Refactoring: Semantic Reconstruction of Domain Models Using LLM Reasoning

Abstract: This study examines the application of large language models (LLMs) for automating domain layer reconstruction in legacy systems, with a specific focus on a case study involving water consumption management. The process begins with a deliberately disordered JSON representation that conflates domain, application, and infrastructure issues. An LLM, specifically GPT-5.2, was employed to identify misplaced methods, inconsistent naming, DTO misuse, incoherent aggregates, and unrelated modules, and subsequently reorganize the model into a structure aligned with Domain-Driven Design (DDD). The structure includes entities, value objects, aggregates, domain services, domain events, and repositories. The methodology involves encoding the legacy model as JSON, applying an LLM-based diagnosis and reconstruction pipeline, and producing both a refined domain model and a categorized catalogue of corrections. A comparative analysis of candidate LLMs, informed by recent code-centric benchmarks, such as SWE-bench and LiveCodeBench, supports the selection of GPT-5.2 as the primary model for this study. The findings indicate that the LLM can swiftly recover key domain concepts and achieve semantically consistent refactoring, a task that typically requires extensive manual effort. This suggests that LLM-assisted domain reconstruction is a promising adjunct to traditional refactoring practices and can facilitate continuous architectural improvements in organizations.

Author 1: Mohamed El BOUKHARI

Author 2: Nassim KHARMOUM

Author 3: Soumia ZITI

Keywords: Domain-Driven Design; large language models; AI-driven software refactoring legacy systems modernization; semantic code analysis; architecture reconstruction; GPT; LLM; domain layer reconstruction; AI-assisted software engineering

Paper 84: Enhancing Telecom Churn Prediction Using Emotion-Driven and Behavioral Engagement Features

Abstract: Accurate churn prediction enables service providers to develop effective retention strategies and promotes revenue stability in the telecommunication industry. This study enhances churn prediction performance by extracting five emotion-driven and behavioral engagement features from a telecom churn dataset. The new features represent derived, experience-oriented indicators constructed from operational usage data rather than direct psychological or survey-based measurements. To assess the effect of these engineered features on predictions, three powerful classifiers (i.e., CatBoost, Random Forest, and XGBoost) were trained and tested in a structured three-stage experimental design. In the first stage, the classifiers were trained and tested using the original dataset (original features only). In the second stage, the original dataset was enriched with five newly derived features (i.e., frustration index, trust score, satisfaction index, service usage score, and international experience index). Finally, in the third stage, only the engineered features were used in the classification process to evaluate their standalone predictive capability. Because the dataset is imbalanced, SMOTE and SMOTE-Tomek were applied to address this issue. The results demonstrate that incorporating these engineered features improves churn prediction performance across the reported evaluation metrics (accuracy, precision, recall, and specificity) for the classifiers and balancing techniques combinations presented. The enriched dataset (original + engineered features) achieves the strongest overall performance compared to using either original features only or engineered features only. Compared to the original features only, the enriched dataset achieved improvements of up to 3.6% in accuracy and 5.8% in recall. These findings indicate that emotion-driven and behavioral engagement features provide meaningful complementary information that enhances churn prediction effectiveness.

Author 1: Huthaifa Aljawazneh

Keywords: Customer churn prediction; emotion-driven features; feature engineering; imbalanced data; classification

Paper 85: Coronary Heart Disease Prediction Using Machine Learning Algorithms

Abstract: Cardiopathy is one of the most serious diseases worldwide with its high morbidity and mortality rates posing a latent risk over time. The objective of this research focuses on evaluating Machine Learning (ML) models such as Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Logistic Regression (LR) for the prediction of coronary heart disease (CHD), with the aim of identifying the most efficient model for this prediction. The model construction followed the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology, which comprises five stages: business understanding, data understanding, data preparation, modeling, and evaluation. The modeling results revealed the superior predictive capability of the XGBoost algorithm for detecting coronary heart disease, compared to Random Forest and Logistic Regression. The assessment of performance metrics (Accuracy, Precision, Sensitivity, and F1 Score) established XGBoost as the reference model, highlighting an F1 Score of approximately 90.8%. This superiority is attributed to its robustness in capturing nonlinear interactions among clinical variables. Consequently, the XGBoost model is selected as the optimal tool for integration into future medical decision support systems. In summary, this ML-based approach provides a highly predictive tool capable of identifying subtle risk patterns from real clinical data. The XGBoost model is a promising candidate for integration into decision support systems and for the optimization of primary prevention protocols for coronary heart disease.

Author 1: Inooc Rubio Paucar

Author 2: Cesar Yactayo-Arias

Author 3: Laberiano Andrade-Arenas

Keywords: Cardiovascular disease; machine learning; prediction; random forest; XGBoost

Paper 86: Integrating Processing-In-Memory into HW/SW Co-Design for Automotive Embedded Systems

Abstract: Continuous expansion of using intelligent learning workloads in modern vehicles, leads Electronic Control Units (ECUs) to manage massive volumes of data while dealing with hard real-time constraints. That said, ECUs must operate under strict power budget due to limited battery capacity and other safety and functional requirements. We study in this research the feasibility of integrating Processing-In-Memory (PIM) approach into the hardware/software co-design process for the emerging ECU architectures. The idea here is to allocate data-centric tasks to in-memory compute units so that computation occurs directly where data is stored. The proposed approach will allow avoiding expensive data traffic, which will improve processing performance and reduce energy consumption. Our work introduces a conceptual framework that reconciles the PIM strengths with automotive requirements and introduces techniques from dynamic voltage scaling to smarter memory management, and is limited to a conceptual architectural proposal without experimental validation or quantitative evaluation. We end up discussing some crucial opportunities and open challenges arising from the implementation of PIM in next-generation automotive systems.

Author 1: Zineb El Kacimi

Author 2: Safae Dahmani

Author 3: Oussama Elissati

Author 4: Mouhcine Chami

Keywords: Electronic Control Units; energy consumption; hardware/software co-design; Processing-In-Memory; real-time constraints

Paper 87: A Feasibility Study of Explainable Machine Learning on Small-Scale Postoperative Voice Data

Abstract: Voice dysfunction is a common complication following thyroid surgery. However, the application of explainable machine learning for predicting postoperative voice recovery remains largely unexplored. Therefore, an investigation was done to examine voice recovery based on acoustic, objective, and glottal features. Voice recordings were collected from female patients before surgery and one month after surgery. Acoustic and glottal parameters, including Quasi Open Quotient, Speed Quotient, age, and others, were automatically extracted from the recordings. Random Forest, Support Vector Machines, and Logistic Regression with Sequential Feature Selection were applied to examine model behavior and identify feature importance. Model stability and interpretability were evaluated across cross-validation folds. Performance metrics varied over folds, highlighting the exploratory and statistically fragile nature of predictions in small datasets. SHAP (SHapley Additive exPlanations) analysis revealed variability in feature contributions, emphasizing the need for cautious interpretation and detailed methodological reporting. Our findings provide preliminary guidance for applying explainable machine learning to small biomedical datasets. They demonstrate the importance of careful methodological design.

Author 1: Noura Haddou

Author 2: Najlae Idrissi

Author 3: Sofia Ben Jebara

Keywords: XAI; explainable AI; SHAP; glottal features; SVM; thyroidectomy; voice recovery

Paper 88: A Dual-Chain and Differential Privacy-Based Solution for Medical Data Privacy Protection and Access Control

Abstract: As living standards rise, people are paying increasing attention to health. Vast quantities of medical data are generated daily, yet each piece contains sensitive information such as patients’ names, mobile numbers, email addresses, and places of employment. Should this information be compromised, the consequences would be irreversible, causing severe damage. Traditional solutions merely implement access control policies, permitting data access only to authorised personnel. While this approach offers some protection, even compliant users cannot be entirely trusted and may engage in malicious activities. Once data is accessed, patients’ sensitive information becomes fully exposed to the user, posing a significant data security risk. Therefore, this study proposes a medical data sharing scheme based on Dual-Chain and differential privacy. It employs a hybrid approach combining private chains, consortium chains, and IPFS. Internal hospital personnel can access data after de-identification, while external parties can only access data that has been de-identified and subsequently augmented with noise. This significantly enhances security. The experimental section of this study also demonstrates that the proposed scheme effectively protects data, while the data shared with external users enables them to successfully complete downstream tasks.

Author 1: Cen Gu

Author 2: Luping Wang

Author 3: Hongjie Wu

Keywords: Blockchain; differential privacy; IPFS; medical data

Paper 89: Leveraging Statistical Invariants to Fortify CNN-Based Medical Diagnostics Against Adversarial Perturbations

Abstract: The integration of artificial intelligence (AI) in medical diagnostics is increasingly jeopardized by adversarial attacks—imperceptible perturbations designed to induce misclassification in Deep Learning models. While Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in medical image analysis, their susceptibility to gradient-based attacks poses a severe risk to patient safety and diagnostic integrity. This study addresses the critical need for robust defense mechanisms in X-ray diagnostics by proposing a Hybrid Ensemble model based on Stacked Generalization. Unlike single-paradigm approaches, our method fuses the spatial feature extraction capabilities of a CNN with the statistical anomaly detection power of a Random Forest (RF). We evaluated this architecture on a curated dataset of X-ray images subjected to Projected Gradient Descent (PGD) attacks with varying perturbation magnitudes (ϵ). The results demonstrate that the Hybrid Ensemble consistently outperforms individual models and standard adversarial training baselines. Under strong attack conditions (ϵ = 0.006), the proposed model achieved an Area Under the Curve (AUC) of 0.919, significantly surpassing the adversarial training baseline (AUC 0.700). Furthermore, the ensemble reduced false positives to 108 compared to 138 for the CNN alone, enhancing clinical reliability. Theoretical motivation for the feature extraction process and extensive experimental validation suggest that leveraging statistical irregularities offers a computationally efficient and robust defense strategy suitable for real-time clinical deployment.

Author 1: Yassine Chahid

Author 2: Anas Chahid

Author 3: Ismail Chahid

Author 4: Aissa Kerkour Elmiad

Keywords: Adversarial defense; medical X-ray; synergistic ensemble; deep learning security; diagnostic robustness; PGD attack

Paper 90: A Transformer-Based Approach for Multimodal Arabic Sentiment Analysis

Abstract: The Multimodal Sentiment Analysis (MSA) land-scape for Arabic content is strikingly underexplored, mainly due to limited datasets and a lack of robust integration methods across text, audio, and image. While transformer-based models like MarBERT and ArBERT achieve strong results on Arabic text, most research remains unimodal and does not fully exploit multimodal synergy. In this work, we propose a three-fold approach for Arabic MSA. First, we finetune robust transformers for each modality, namely ViT, MarBERT, and HuBert for image, Text, and Audio, respectively. Second, we perform an early feature fusion. Third, we use classifiers for sentiment prediction. On the recent Ar-MuSA benchmark released on 2025, our tri-modal fusion system, achieves state-of-the-art performance (F1=0.7756, Accuracy=0.7759), significantly exceeding the multimodal models benchmarked on the Ar-MuSa dataset, as well as the unimodal and bimodal methods. This demonstrates that comprehensive tri-modal fusion and thoughtful classifier selection are essential for accurate, human-centric Arabic sentiment analysis.

Author 1: Ayoub BEN CHEIKHI

Author 2: EL Habib NFAOUI

Keywords: Arabic sentiment analysis; multimodal learning; feature fusion; machine learning; early fusion

Paper 91: HQ-RTVF: High-Quality Real-Time Virtual Try-On Fitting for Diverse Clothing and Body Morphologies

Abstract: The ability to virtually try on clothing items has become an increasingly important feature for e-commerce and online shopping experiences. Real-time virtual try-on remains challenging because existing methods force a trade-off between speed and quality GAN-based approaches achieve high visual fidelity but at low frame rates, while faster methods sacrifice realism. HQ-RTVF is a diffusion-based framework that resolves this trade-off through three architectural innovations: running the diffusion U-Net entirely in the VAE’s compressed latent space (64×64×4 instead of 512×512×3), limiting denoising to 20 steps with FP16 mixed-precision computation, and parallelizing pose estimation and garment encoding to eliminate sequential bottlenecks. The system uses DensePose and DeepLabv3+ for body pose and segmentation, a CLIP-based garment encoder for fine-grained fabric representation, and an attention-guided fusion decoder that maintains temporal coherence across video frames— distinguishing it from static image methods like VITON-HD and HR-VITON. An adaptive masking mechanism handles diverse garment types from cropped tops to full-length dresses. Evaluated on VITON-HD and DressCode datasets, HQ-RTVF achieves SSIM of 0.950 and LPIPS of 0.067, while operating in real-time with only 4.2 GB GPU memory.

Author 1: Ilham KACHBAL

Author 2: Khadija Arhid

Author 3: Said El Abdellaoui

Keywords: Virtual try-on; diffusion models; real-time processing; deep learning; garment synthesis; pose estimation

Paper 92: Density-Guided Adaptive Patch Learning for Robust Crowd Counting

Abstract: Accurate crowd counting in real-world scenes re-mains challenging due to severe occlusions, perspective distortion, and large intra-scene density variation. Recent deep learning based approaches typically address these challenges using patch-level learning, where images are divided into fixed grids or randomly cropped patches. These approaches then estimate the count in each patch. However, such fixed partitioning strategies often fail to align with the irregular spatial distribution of crowds. This leads to heterogeneous density patterns within patches where the models fail to produce the accurate count. In this study, we propose a simple, yet effective Density-Guided Adaptive Patch Learning framework for crowd counting. Instead of relying on fixed-size patches, we first obtain a coarse density estimation to capture the global density structure of a scene. Based on this estimate, the image is dynamically partitioned into density-homogeneous regions, where dense areas are represented using smaller patches and sparse regions using larger patches. Each adaptive patch is then processed independently for density estimation, and the resulting predictions are fused to produce the final crowd density map. The proposed framework is model-agnostic and can be seamlessly integrated with existing crowd counting networks without architectural modification. Extensive experiments on benchmark datasets demonstrate that the pro-posed adaptive partitioning consistently achieves lower Mean Absolute Error (MAE) in counting accuracy and localization compared to fixed patch-based baselines, particularly in scenes with strong density variation.

Author 1: Abdullah N Alhawsawi

Keywords: Computer vision; deep learning; crowd counting

Paper 93: Forward Selection for Time Series-Based Qubit Generation via Parameterized Quantum Gates

Abstract: Quantum data processing requires classical data to be encoded into quantum states. Current noisy intermediate-scale quantum devices have a limited number of qubits that are stable only briefly. Encoding classical data into qubits is the initial step in Quantum Machine Learning (QML), and effective encoding is crucial for quantum processing. This algorithms for data processing are still emerging, and compact data representations are essential for their success. This research proposes a novel data encoding technique using uniformly controlled rotation gates, achieving high storage density by encoding real-valued time series data as qubit rotations. The model uses a binary representation for computations on time series data, reducing the number of quantum measurements needed. The research explores quantum forward propagation in simulations to improve prediction accuracy for time series signals using parameterized quantum circuits, handling trends, noise, and sinusoidal components. The efficiency of the encoding process depends on data volume and chosen encoding, with potential infinite loading time in the worst case. This study presents a Forward Selection Time Series Data Pro-cessing and Feature Extraction Model for Qubits generation with Parameterized Quantum Gates (FSDPFEM-PQG), demonstrating superior performance in quantum representations compared to existing models.

Author 1: Singaraju Srinivasulu

Author 2: Nagarajan G

Keywords: Quantum bits; Quantum Machine Learning; quantum algorithms; quantum measurements; Parameterized Quantum Gates; feature extraction; time-series data

Paper 94: A Hybrid DMD–TCN Framework for Interpretable Short-Horizon Prediction of 6-DOF Ship Motions

Abstract: Accurate short-horizon prediction of six degrees of freedom (6-DOF) vessel motions is essential for autonomous navigation, motion compensation, and operational decision making. Traditional seakeeping models rely on hydrodynamic coefficients that are seldom available for full-scale vessels, while purely data-driven approaches may struggle to maintain physical consistency. This study introduces a hybrid physics–machine learning framework that combines Dynamic Mode Decomposition (DMD), which approximates the vessel’s dominant linear drift dynamics, with a causal Temporal Convolutional Network (TCN) that learns nonlinear residual corrections from a 12-hour historical window of environmental, geometric, and motion features. DMD provides an interpretable surrogate of the vessel dynamics through its eigenvalues, growth rates, and mode shapes, serving as a data-derived linear transfer operator. The TCN predicts only the residual departure from this structured baseline, ensuring a stable and causal forecasting architecture. Evaluation on full-scale field data shows that the hybrid model improves prediction accuracy for heave and achieves performance comparable to DMD for surge, while underperforming in sway, roll, pitch, and yaw due to the limited observability of key physical drivers at hourly resolution. These results highlight both the strengths and limitations of residual learning when important nonlinear forcing mechanisms and control inputs are unmeasured. Overall, the study demonstrates that hybrid physics–machine learning approaches provide valuable interpretability and diagnostic insight, even when data limitations are constrained. The framework offers a principled foundation for incorporating additional physical inputs, higher-frequency measurements, and physics-informed architectures in future work on operational ship-motion forecasting.

Author 1: Enock Tafadzwa Chekure

Author 2: Kumeshan Reddy

Author 3: John Fernandes

Keywords: Dynamic Mode Decomposition; Temporal Convolutional Network; hybrid learning; seakeeping and vessel response; data-driven modelling; Koopman operator methods

Paper 95: SecureDML:An Intelligent Framework for Preventing Poisoning Attacks in Distributed Machine Learning Systems

Abstract: The security and protection of models in dis-tributed machine learning (ML) systems require high emphasis on adversarial threats, including poisoning attacks. This study contains a complete framework that integrates different advanced techniques to monitor poison attacks and prevent such attacks for the effective functioning of machine learning systems. The proposed system integrates hybrid encryption for security, and a subsequent anomaly detection method using autoencoders. SHapley Additive exPlanations-based interpretability method is used to enhance model transparency. Hybrid encryption combines the RSA and AES methods to keep data and model parameters secret, and autoencoders provide effective identification of poisoning attack patterns through abnormal data observations. This method is implemented using multimodal datasets such as CIFAR 100 and AG News datasets. Finally,the effectiveness of this method can be evaluated using confusion matrix, comparison graphs. It works as a comprehensive solution that benefits various ML applications, such as healthcare, autonomous vehicles, Large Language Models, etc., for enhancing security along with integrity protection.

Author 1: Archa A. T

Author 2: Kartheeban K

Keywords: Poisoning attacks; SHapley Additive exPlanations; anomaly detection; federated learning; autoencoders

Paper 96: Post-Quantum Module Learning with Rounding-Based Public Key Encryption Using Incomplete Number Theoretic Transform

Abstract: Post Quantum Cryptographic (PQC) techniques are widely used in encryption standards and digital signatures. Lattice-based post-quantum cryptographic techniques have been reported in the last decades. The present work proposes an optimized quantum-safe lattice public key encryption (PKE) scheme based on the Module Learning with Rounding (MLWR) problem, enhanced by the use of Incomplete Number Theoretic Transform (NTT). The objective of the proposed scheme is to achieve efficient encryption and decryption while maintaining robust security in accordance with the National Institute of Standards and Technology (NIST) recommendations. The incomplete NTT relaxes the modulus q requirement, enables a smaller modulus for efficient arithmetic, and reduces computational complexity. This approach results in significant improvements in the speed of key generation, encryption, and decryption with a marked reduction in rejection probability, compared to schemes utilizing complete NTT. The proposed scheme demonstrates competitive performance against other lattice-based encryption schemes such as Kyber and Frodo. It shows lower encryption and decryption times while offering comparable security levels. Proposed scheme is at least a hundred times faster than Frodo lattice-based public-key encryption schemes. For NIST-recommended security level, in proposed scheme, each encryption needs an average of 300K CPU cycles, and each decryption needs 120K CPU cycles. Additionally, modulus 7937 enables a reduction in key and ciphertext sizes, optimizing the scheme for practical deployment in resource-constrained environments. Performance evaluations confirm the practicality of the scheme with substantial reductions in computational overhead, making it a highly efficient and secure candidate for post-quantum encryption.

Author 1: Anupama Arjun Pandit

Author 2: Arun Mishra

Keywords: Post-quantum cryptography; public key encryption; lattice-based cryptography; Incomplete Number Theoretic Transform; Module Learning with Rounding; Learning with Errors

Paper 97: Abnormal State Detection of Industrial Tools Based on the MGC-YOLOv8 Algorithm

Abstract: As intelligent manufacturing advances toward precision and automation, cutting tool condition critically impacts product quality, equipment safety, and production efficiency. Anomalies like wear, chipping, or fracture cause workpiece scrapping and machine failure, demanding efficient online monitoring. Traditional manual or image-based methods suffer from low accuracy in complex environments. Although deep learning excels in industrial defect detection, existing end-to-end detectors exhibit insufficient recall and localization precision for millimeter-scale cracks and blurred tool boundaries. To address these challenges, we propose MGC-YOLOv8, an enhanced framework built upon the YOLOv8 backbone. A Multi-Scale Edge-Dual Fusion (MSEDF) module is introduced to integrate feature maps across different scales, thereby strengthening the detection of minor defects. Furthermore, a Global-to-Local Spatial Aggregation (GLSA) module enriches feature representations by simultaneously capturing global context and local details. A Convolutional Block Attention (CBAM) module is embedded upstream of the prediction head to adaptively highlight critical features in both channel and spatial dimensions. Although the integration of MSEDF, GLSA, and CBAM introduces a marginal runtime overhead and a slight increase in parameter count, the optimized architecture preserves real-time inference speeds that fully satisfy the requirements of industrial inspection systems. Experimental results demonstrate that MGC-YOLOv8 substantially outperforms the baseline YOLOv8n, achieving 88.1% precision, 87.9% recall, 92.5% mAP@0.5 and 69.6% mAP@0.5:0.95 on our test set.

Author 1: Guan Yang

Author 2: Xiang Cheng

Author 3: Miao Wang

Author 4: Ziyue Huang

Author 5: Hao Tang

Author 6: Yujun Chen

Keywords: Object detection; surface defect detection; YOLOv8

Paper 98: Tomato Maturity Analysis: A Comparative Study of Detection and Instance Segmentation Using YOLOv8

Abstract: The accurate visual analysis of fruit maturity in complex agricultural scenes remains a fundamental challenge due to gradual appearance changes, object overlap, and partial occlusion. This study addresses tomato maturity analysis, formally defined as instance-level binary classification and spatial localization under varying degrees of visual density. While bounding-box-based object detection is widely used, it often lacks precision in dense clusters. We present a controlled experimental comparison between object detection and instance segmentation using a common YOLOv8-medium (YOLOv8m) backbone to isolate the effect of spatial representation. Experimental results demonstrate that instance segmentation achieves superior localization accuracy and boundary consistency, reaching a mask-based mAP@0.5:0.95 of 0.817. These findings suggest that pixel-level supervision effectively reduces localization ambiguity, providing a robust foundation for automated agricultural monitoring.

Author 1: Salma Ait Oussous

Author 2: Rachid El Bouayadi

Author 3: Driss Zejli

Author 4: Aouatif Amine

Keywords: Tomato maturity detection; computer vision; object detection; instance segmentation; image analysis; Deep Learning

Paper 99: Time Series Anomaly Detection Based on Entropy-Sparsified Time-Frequency Fusion and MsRwGWO Meta-Optimization

Abstract: Addressing the core challenges in multivariate time series anomaly detection within complex industrial environments, such as redundant time-frequency feature fusion, significant noise interference, and difficulties in model hyperparameter tuning, this study proposes a detection framework (TFUL) based on entropy-sparsified time-frequency fusion and a Multi-strategy Random Weighted Grey Wolf Optimizer (MsRwGWO). The main contributions of this work include: 1) A dual-domain entropy sparsification fusion mechanism is designed, which dynamically evaluates and filters crucial temporal segments and frequency components via information entropy, enabling adaptive and redundancy-resistant feature fusion. 2) A heterogeneously collaborative feature extraction network is constructed. The temporal branch, SoftShapeNet, integrates multi-scale convolutions and a Mixture of Experts (MoE) to capture local polymorphic shapes, while the frequency branch, FrequencyDomainProcessor, employs a learnable Mahalanobis distance to model nonlinear spectral dependencies among channels, surpassing the limitations of fixed transformations. 3) The MsRwGWO meta-optimization strategy is proposed, which incorporates dynamic weighting and multi-strategy perturbation mechanisms, significantly enhancing the efficiency and quality of hyperparameter search. Experiments conducted on several public datasets demonstrate that the pro-posed method outperforms mainstream comparative models in terms of detection accuracy and robustness, providing an effective solution for industrial time series anomaly detection.

Author 1: Xiaogang Yuan

Author 2: Jiaxi Chen

Author 3: Dezhi An

Author 4: Jianxin Wan

Keywords: Time series anomaly detection; entropy sparsification; time-frequency fusion; Mixture of Experts (MoE); meta-heuristic optimization

Paper 100: Blockchain-Based Multi-Chain Data Supervision Mechanism for Traditional Chinese Medicine Traceability System

Abstract: Addressing the challenges of Traditional Chinese Medicine (TCM) traceability systems, including heavy data storage burdens, poor privacy protection, and susceptibility to tampering, this study establishes a highly secure and trustworthy traceability supervision system for the entire Chinese medicine supply chain, which enhances product quality and safety assurance. Centred on the Hyperledger Fabric consortium blockchain as its core architecture, a multi-chain integration framework comprising one regulatory main chain plus five organisational sub-chains is proposed to achieve permission control, data isolation, and privacy. A multi-mode encrypted data storage mechanism is designed, integrating China’s national cryptographic algorithms SM4 and SM3 with CP-ABE attribute-based encryption to enable tiered management of private and non-private data. Zero-knowledge proof technology safeguards identity privacy during cross-chain data transmission, while QR codes and environmental data collection mechanisms enhance data entry efficiency and authenticity. The system achieves end-to-end traceability from cultivation and processing through transportation, warehousing, and sales. Comparative performance analysis shows that the proposed framework effectively alleviates data storage pressure, ensures data validity, enhances data security, and improves collaborative efficiency among organizations across the TCM supply chain. The proposed multi-chain integrated Chinese medicine traceability and supervision system enables efficient collaboration and trustworthy traceability across the entire Chinese medicine industry chain, while safeguarding data security and privacy, and has significant application and promotion value. Future integration with artificial intelligence and big data technologies could further enhance the system’s intelligent analysis and decision-support capabilities.

Author 1: Rongjun Chen

Author 2: Yun Sun

Author 3: Feng Xue

Author 4: Yongzhi Ma

Author 5: Xinyu Wu

Author 6: Xianxian Zeng

Author 7: Jiawen Li

Author 8: Jinchang Ren

Keywords: Blockchain; traceability; multi-chain architecture; Hyperledger Fabric; Traditional Chinese Medicine

Paper 101: Attention-Guided Fusion of EfficientNet-B0 and Swin Transformer for Cervical Cancer Classification

Abstract: The interpretation of colposcopy images is a critical yet subjective component of cervical cancer screening. To enhance this process, we propose a novel hybrid deep learning framework for the classification of cervical lesions. Our model integrates EfficientNet-B0, adept at extracting localized hierarchical features, with a Swin-Tiny Transformer, which excels at modeling long-range dependencies and global context. Moving beyond basic fusion techniques, we introduce a novel cross-attention fusion mechanism, augmented with channel and spatial attention modules. This design selectively highlights the most discriminative inter-feature relationships while maintaining computational efficiency. Evaluated on the International Agency for Research on Cancer (IARC) colposcopy image dataset, our framework achieves an accuracy of 94.76%, significantly outperforming a concatenation-based fusion model (83.99%). This represents an absolute improvement of 10.77 percentage points and captures 67.3% of the residual performance margin toward perfect ac-curacy. The model also demonstrates robust performance across other metrics, including a precision of 94.68%, recall of 94.82%, F1-score of 94.74%, and a Cohen’s Kappa of 89.48%. These results indicate that our approach can enhance both the accuracy and reliability of cervical cancer screening, offering valuable support for clinical decision-making.

Author 1: Twisibile Mwalughali

Author 2: Emmanuel C. OGU

Author 3: Evason Karanja

Keywords: Cervical cancer classification; deep learning model; colposcopy; cross-attention fusion; EfficientNet; Swin Transformer

Paper 102: Optimizing Resource Allocation for Crisis-Resilient Healthcare Robotics: An Integrated MLR, MDP, and Petri Net Approach

Abstract: Global crises, such as pandemics and climate-related disasters, place unprecedented strain on healthcare Systems, exposing weaknesses in resource management and patient care. This study aims to address these challenges by developing an integrated computational framework for crisis-resilient healthcare robotics. We propose a unified approach that combines Multiple Linear Regression (MLR), Markov Decision Processes (MDPs), and Petri Nets. MLR is applied to predict the Average Length of Stay (ALOS) using patient and hospital data. These forecasts inform MDPs, which guide admission and triage decisions under uncertainty. Petri Nets are employed to model and validate patient flow and hospital workflows, ensuring feasibility and efficiency. Case studies, including ICU bed prioritization and disaster logistics, demonstrate that the proposed framework improves adaptability and resource utilization while supporting structured decision guidance. Simulation results highlight enhanced system efficiency, better patient prioritization, and reduced congestion during surge conditions. The integration of predictive analytics, probabilistic optimization, and workflow modeling provides a robust decision-support system for health-care robotics in crisis scenarios. This interdisciplinary framework offers practical solutions for improving resilience, scalability, and patient outcomes, providing a structured foundation for enhancing resilience and coordination in healthcare systems facing future emergencies.

Author 1: Ikram Dahamou

Author 2: Ayoub Elbazzazi

Author 3: Cherki Daoui

Keywords: Healthcare robotics; Markov Decision Processes; Petri Nets; Multiple Linear Regression; crisis management; ICU resource allocation

Paper 103: Enhanced Ant Colony Optimization for Capacitated Vehicle Routing Problem with Time Windows in Franchise Distribution

Abstract: Efficient routing for distributing goods to multiple franchisee locations requires optimization techniques capable of handling vehicle capacity limits, heterogeneous time windows, and operational constraints, making conventional brute-force or map-based approaches infeasible due to the NP-hard nature of the problem. This study presents an enhanced Ant Colony Optimization (ACO) algorithm for solving the Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) in a franchisor–franchisee logistics setting. The proposed enhancement incorporates feasibility filtering to enforce capacity and time-window constraints during route construction and adaptive pheromone updating to improve convergence stability. Using real franchisee coordinates, demand values, and operational time windows, the experiments configured with α = 2, β = 1, ρ = 0.05, and a 150-iteration limit demonstrate that the enhanced ACO achieves a minimum total route distance of 46.90 km with zero variance across 10 simulations, indicating highly stable convergence. Comparative evaluation shows that the enhanced ACO improves route efficiency by 11.4% compared to standard ACO and 15.2% relative to a representative Genetic Algorithm baseline. Implemented in a web-based environment using JavaScript for visualization and Java for computation, the approach provides a practical decision-support tool for Indonesian franchise logistics. The algorithm exhibits an observed computational complexity of θ(n4), making it suitable for small to medium-scale distribution networks involving strict delivery time windows.

Author 1: Dian Rachmawati

Author 2: Tommy Lohil

Author 3: Jos Timanta Tarigan

Keywords: Ant Colony Optimization; CVRPTW; heuristics; distribution routing; logistics optimization

Paper 104: Designing an Attack-Vector-Based Taxonomy for IoT Malware

Abstract: This study presents a literature-derived, attack-vector-based taxonomy for IoT malware and complements it with an empirical validation using supervised machine learning. Building on prior surveys and taxonomies of IoT security and malware behavior, we synthesize how existing studies implicitly or explicitly describe infection vectors such as credential abuse, exposed services, firmware exploitation, internal lateral movement, and supply-chain compromise. The resulting taxonomy organises IoT malware according to initial entry mechanisms rather than post-compromise capabilities, providing a vector-centric perspective that aligns more naturally with risk assessment and defensive planning. To demonstrate the practical relevance of this taxonomy, we implement a supervised malware detection model operating on Windows Portable Executable (PE) files. Using malware samples collected from public repositories (e.g., VirusShare and MalwareBazaar) and benign executables from open-source projects, we extract structural, statistical, and metadata-based PE features and train an Extreme Gradient Boosting (XGBoost) classifier with Synthetic Minority Over-sampling Technique (SMOTE) for class balancing. The model achieves an accuracy of 98.13% with balanced F1-scores for both malware and benign classes, illustrating that feature-engineered supervised models can effectively support taxonomy-informed detection strategies. The combined conceptual and empirical view highlights how attack-vector taxonomies, IoT threat modeling, and machine learning-based detection can be jointly leveraged to strengthen IoT cyber defense.

Author 1: Huda Aldawghan

Author 2: Mounir Frikha

Keywords: IoT security; malware taxonomy; attack vectors; cyber threat intelligence; network defense

Paper 105: Enhancing Malware Detection Using Machine Learning Models on Static Features

Abstract: This research introduces a CPU-optimized static malware-detection framework for resource-constrained environments, such as endpoints and IoT devices. We address the significant challenge of high memory and computational demands by proposing a robust, memory-safe data ingestion pipeline. This pipeline exclusively extracts histogram-based static features, employs type compression, and utilizes batch-wise loading with global sample limits to prevent memory overflows on systems with only 16 GB of RAM and no GPU support. Our core contribution is a compact stacking ensemble composed of three high-efficiency gradient-boosting models: LightGBM, CatBoost, and XGBoost, with a LightGBM meta-learner. This novel ensemble structure enables efficient, CPU-only training and inference while ensuring strong detection performance. Evaluated on the EMBER 2024 dataset, the framework achieves 86.99% accuracy, 0.87 F1-score, and 0.9473 AUC. This work fills a critical gap by demonstrating that carefully optimized gradient-boosting ensembles can serve as a highly deployable alternative to resource-intensive Deep Learning methods in limited security situations.

Author 1: Ashwag Alotaibi

Author 2: Mounir Frikha

Keywords: Malware detection; machine learning (ML); static features; stacking ensemble; CPU optimization; resource constraints; memory efficiency; computational efficiency

Paper 106: Stability-Aware QUBO Feature Selection for Tabular Classification Under Repeated Nested Cross-Validation

Abstract: Quadratic Unconstrained Binary Optimization (QUBO) provides a principled framework for feature selection by encoding relevance–redundancy trade-offs and explicit constraints directly in a combinatorial objective. This study presents a stability-aware QUBO pipeline for tabular binary classification, evaluated on two standard benchmarks, namely Breast Cancer Wisconsin Diagnostic (569 samples, 30 features) and Pima Indians Diabetes (768 samples, 8 features; clinically invalid zeros treated as missing and imputed within folds). We study four QUBO variants spanning a base relevance–redundancy formulation, an exact-cardinality formulation enforcing a fixed budget k, a stability-regularized formulation that incorporates bootstrap uncertainty estimates of relevance and redundancy directly into the QUBO objective, and a performance-weighted relevance variant based on inner-CV univariate utility. All methods are assessed under repeated nested stratified cross-validation (5 outer folds × 3 repeats, n = 15 outer test evaluations), reporting AUC-ROC, AUC-PR, MCC, and Brier score with 95% confidence intervals, alongside selection stability via mean Jaccard similarity across outer-fold selected subsets. Results show that QUBO-based selection is competitive with strong classical baselines (RFECV, L1-logistic, permutation-importance ranking, and mutual information) while enabling strict budget control and transparent stability diagnostics. On the near-ceiling Breast Cancer benchmark, predictive differences are marginal and the main differentiators become subset-size control and stability; on Pima, QUBO-k remains competitive while enforcing strict cardinality constraints. These findings support QUBO as a practical framework when budgeted, interpretable, and reproducible feature selection is required, though evaluation is limited to low-dimensional tabular settings.

Author 1: Marco Fidel Mayta Quispe

Author 2: Leonid Alemán Gonzales

Author 3: Charles Ignacio Mendoza Mollocondo

Author 4: Nayer Tumi Figueroa

Author 5: Juan Carlos Juarez Vargas

Author 6: Godofredo Quispe Mamani

Keywords: Feature selection; QUBO; simulated annealing; nested cross-validation; selection stability; Jaccard similarity; probability calibration; tabular classification

Paper 107: Attention-Guided Bidirectional Temporal Modelling with Graph-Based Regional Spatial Context for Bajra Crop Yield Prediction

Abstract: Bajra (pearl millet) is a very important crop in Rajasthan, India, since it is drought-resistant, nutritious, and culturally important. But its productivity is becoming vulnerable to changes in climate, such as erratic rain and temperature changes, and thus precise estimation of yield is vital. Crop Yield Prediction (CYP) indicators like soil decomposition, rainfall and meteorological patterns are slowly evolving, exhibiting long-term temporal dependency and propagating over time. Conventional cropping prediction algorithms based on artificial intelligence process the historical data and these indicators in a unidirectional manner. While mapping the temporal dependencies, these algorithms consider each year independently and do not capture the delayed effect, like salt degradation. To address this issue, the study proposes a region-based spatiotemporal model with an attention-guided Bidirectional LSTM (Long-Short Term Memory) framework for CYP, termed as G-BiLSTM. The proposed model reproduces the spatial relationships between districts via GCN (Graph Convolution Network) -based immediate neighbour extraction. Further, a Bidirectional LSTM is used to model multi-year CYP temporal features, allowing each annual observation to be encoded using both past and future temporal context. A variance-reduced and comprehensible representation is produced by integrating an attention mechanism to adaptively highlight the most informative years within a temporal window. Using 15 agroenvironmental characteristics, including understudied elements like saline and alkaline soil composition, the framework is assessed on a dataset that includes 32 districts in Rajasthan over 13 years (2007–2019). The suggested attention-enhanced BiLSTM consistently outperforms traditional temporal models, achieving lower prediction error and better generalisation, according to experimental results analysis using a three-year sliding temporal window. For regional crop yield forecasting, the suggested method offers a scalable solution.

Author 1: Mamta Kumari

Author 2: Suman

Author 3: Devendra Prasad

Keywords: Bajra crop yield prediction; regional crop yield fore-casting; attention-guided Bidirectional LSTM; saline and alkaline soil composition

Paper 108: Scalability of Predictive Models on Multi-Core CPUs and GPUs: An Empirical Analysis

Abstract: The scalability of predictive models has become a critical factor in modern machine learning, as data volumes grow and computational resources diversify. This study presents an empirical benchmark of three widely used regression paradigms: Elastic Net, XGBoost, and Multi-Layer Perceptrons (MLPs). The Obesity Estimation dataset is used to evaluate both predictive performance and computational scalability across multi-core CPUs and GPUs. Unlike prior studies that primarily emphasize accuracy, we explicitly examine the trade-offs between accuracy, training time, and hardware efficiency. Models are evaluated under staged training loads (10–100% of data) with grid-searched hyperparameters (for Elastic Net and XGBoost) and regularized deep architectures (for MLP). Results demonstrate that while XGBoost achieves the highest predictive accuracy (R2 = 0.91), it incurs significant computational overhead on CPUs, whereas GPU acceleration substantially improves its scalability. MLPs provide competitive accuracy (R2 = 0.87) with an order-of-magnitude lower training time on GPUs, making them attractive for rapid or repeated retraining. Elastic Net offers interpretability and linear scalability on CPUs, but lags in predictive power. These findings provide practitioners with a decision framework: XGBoost for maximum accuracy, MLPs for efficient retraining, and Elastic Net for interpretability and small-scale tasks. More broadly, this work highlights that hardware selection is as important as algorithm choice, with GPUs serving as enablers of state-of-the-art performance on structured data.

Author 1: Atif Mahmood

Author 2: Wan Joe Dean

Author 3: P. Ganesh Kumar

Author 4: Adnan N. Qureshi

Keywords: Scalability; XGBoost; Neural Networks; Elastic Net; resource efficiency