IJACSA Volume 16 Issue 5 - thesai.org

IJACSA Volume 16 Issue 5

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: Enhancing Federated Learning Security with a Defense Framework Against Adversarial Attacks in Privacy-Sensitive Healthcare Applications

Abstract: Federated learning (FL) is a cutting-edge method of collaborative machine learning that lets organizations or companies train models without exchanging personal information. Adversarial attacks such as data poisoning, model poisoning, backdoor attacks, and man-in-the-middle attacks could compromise its accuracy and reliability. Ensuring resistance against such risks is crucial as FL gets headway in fields like healthcare, where disease prediction and data privacy are essential. Federated systems lack strong defenses, even though centralized machine learning security has been extensively researched. To secure clients and servers, this research creates a framework for identifying and thwarting adversarial attacks in FL. Using PyTorch, the study evaluates the framework’s effectiveness. The baseline FL system achieved an average accuracy of 90.07%, with precision, recall, and F1-scores around 0.9007 to 0.9008, and AUC values of 0.95 to 0.96 under benign conditions. With AUC values of 0.93 to 0.94, the defense-enhanced FL system showed remarkable resilience and maintained dependable classification (precision, recall, F1-scores ~0.8590–0.8598), despite a 4.1% accuracy decline to 85.97% owing to security overhead. With an 84.33% attack detection rate, 99.32% precision, 96.62% accuracy and a low false positive rate of 0.15%, the defense architecture performed exceptionally well in adversarial attacks. Trade-offs were identified via latency analysis: the defense-enhanced system stabilized at 54 to 56 seconds, while the baseline system averaged 13-second rounds. With practical implications for safe, robust machine learning partnerships, these findings demonstrate a balance between accuracy, efficiency and security, establishing the defense-enhanced FL system as a reliable option for privacy-sensitive healthcare applications.

Author 1: Frederick Ayensu

Author 2: Claude Turner

Author 3: Isaac Osunmakinde

Keywords: Federated learning; machine learning; privacy; adversarial attacks; defense framework; global model; healthcare; disease prediction

Paper 2: Automated Analysis of Glucose Response Patterns in Type 1 Diabetes Using Machine Learning and Computer Vision

Abstract: This study presents an automated and data-driven framework for analysing glucose response patterns in individuals with Type 1 diabetes by integrating machine learning and computer vision methodologies. The system leverages multimodal data inputs, including food images, continuous glucose monitoring (CGM) data, and time-series meal logs to model glycaemic variability and infer personalized dietary effects. Using a dataset comprising over eighty annotated meals from eight subjects, the framework extracts nutritional features from food images via convolutional neural networks (CNNs) with attention mechanisms and correlates them with postprandial glucose trajectories. The analysis reveals substantial inter-individual variability and identifies critical temporal and nutritional factors influencing glucose dynamics. Results demonstrate the system’s capability to detect patterns predictive of glycemic responses, enabling the development of tailored dietary recommendations. This approach offers a scalable tool for personalized diabetes management and paves the way for future integration into real-time decision support systems.

Author 1: Arjun Jaggi

Author 2: Aditya Karnam Gururaj Rao

Author 3: Sonam Naidu

Author 4: Vijay Mane

Author 5: Siddharth Bhorge

Author 6: Medha Wyawahare

Keywords: Continuous glucose monitoring; glucose response; Type 1 diabetes; food image analysis; dietary pattern recognition; time-series analysis

Paper 3: Quantized Object Detection for Real-Time Inference on Embedded GPU Architectures

Abstract: Deploying deep learning-based object detection models like YOLOv4 on resource-constrained embedded ar-chitectures presents several challenges, particularly regarding computing performance, memory usage, and energy consumption. This study examines the quantization of the YOLOv4 model to facilitate real-time inference on lightweight edge devices, focusing on NVIDIA’s Jetson Nano and AGX. We utilize post-training quantization techniques to reduce both model size and computational complexity, all while striving to maintain acceptable detection accuracy. Experimental results indicate that an 8-bit quantized YOLOv4 model can achieve near real-time performance with minimal accuracy loss. This makes it well-suited for embedded applications such as autonomous navigation. Additionally, this research highlights the trade-offs between model compression and detection performance, proposing an optimization method tailored to the hardware constraints of embedded architectures.

Author 1: Fatima Zahra Guerrouj

Author 2: Sergio Rodriiguez Florez

Author 3: Abdelhafid El Ouardi

Author 4: Mohamed Abouzahir

Author 5: Mustapha Ramzi

Keywords: Object detection model; quantization; embedded architectures; real-time

Paper 4: End-to-End Current Consumption Estimation for a Driving System of a Mobile Robot Considering Geology

Abstract: Mobile robots are often tasked with environmental surveys and disaster response operations. Accurately estimating the energy consumption of these robots during such tasks is essen-tial. Among the various components, the drive system consumes the most energy and exhibits the greatest fluctuations. Since these energy fluctuations stem from variations in current consumption, it is crucial to estimate the drive system’s current consumption with high accuracy. However, existing research faces challenges in accurately estimating current consumption, particularly when the ground geology changes or when internal states cannot be measured. Moreover, there is no clearly defined methodology for estimating the current consumption of a mobile robot’s drive system under unknown geological conditions or internal states. To address this gap, the present study aims to develop an end-to-end method for estimating the current consumption of a mobile robot’s drive system, taking ground geology into consideration. To achieve this, we propose a novel approach for collecting interaction data and generating a current consumption model. For data collection, we introduce a method that effectively captures the internal and external factors influencing the drive system’s current consumption, as well as their interactions. This is accomplished by treating the physical phenomena resulting from the interaction between the driving mechanism and the ground as vibrations. Additionally, we propose a method for generating a current consumption model using a neural network, accounting for measurement errors, outliers, noise, and global current fluctuations. The effectiveness of the proposed method is demonstrated through experiments conducted on three different ground types using a skid-steering mobile robot.

Author 1: Shota Chikushi

Author 2: Yonghoon Ji

Author 3: Hanwool Woo

Author 4: Hitoshi Kono

Keywords: Current consumption estimation; mobile robot; neu-ral network; snow environment

Paper 5: Enhancing Industrial Cybersecurity with Virtual Lab Simulations

Abstract: The increasing integration of Industrial Control Systems (ICS) within production environments underscores the urgent need for robust cybersecurity measures. However, securing these devices without disrupting ongoing operations presents a significant challenge. This study introduces a virtual laboratory environment that simulates real-world ICS networks, including a misconfigured Active Directory (AD) domain and a Supervisory Control and Data Acquisition (SCADA) node, to train cybersecurity professionals in recognizing and mitigating vulnerabilities. We propose a comprehensive setup of virtual machines—Windows Server, Windows Workstations, and Kali Linux—and follow the Purdue model for network segmentation, effectively bridging theory with hands-on practice. Demonstrating various penetration testing tools (e.g., Impacket, Kerbrute, Chisel, Socat, and TeslaCrypt ransomware), this work reveals how a single misconfiguration, such as disabling Kerberos pre-authentication, can cascade into severe breaches, including ransomware attacks on critical devices. Our preliminary results show that the virtual laboratory approach strengthens business continuity and resilience by enabling real-time testing of countermeasures without risking production downtime. This ongoing research aims to provide a practical, adaptable, and standards-aligned solution for cybersecurity training and threat response in industrial setting.

Author 1: Hamza Hmiddouch

Author 2: Antonio Villafranca

Author 3: Raul Castro

Author 4: Volodymyr Dubetskyy

Author 5: Maria-Dolores Cano

Keywords: Cybersecurity; industrial control system; ransomware; virtual lab

Paper 6: HCAT: Advancing Unstructured Healthcare Data Analysis Through Hierarchical and Context-Aware Mechanisms

Abstract: To that end, this study presents the Hierarchical Context-Aware Transformer (HCAT), a new model to perform analysis on unstructured healthcare data that resolves significant problems related to medical text. In the proposed model, the hierarchical structure of the system is integrated with the context-sensitive mechanisms to process the healthcare documents at sentence level and document levels. HCAT complies with domain knowledge by a specific attention module and uses a detailed loss function that focuses on classification accuracy besides encouraging domain adaptation. The quantitative experiment shows that HCAT is a better choice than Bi-LSTM and BERT for sentence representation. The model attains 92.30% test accuracy on medical text classification, conversing with high computational efficiency; batch processing time is about 150ms, while the memory consumed is 320 MB. The proposed architecture for clinical text representation facilitates the incorporation of long-range dependencies for clinical story representation, whereas the context-sensitive layer supports a better understanding of medical language. Precision and recall are significant because of the healthcare application of the model; the model has an accuracy of 91.8% and a recall of 93.2%. From these results, it can be concluded that HCAT presented significant progress in computing healthcare data. It provides a highly practical application for real-world extraction of medical data from unformatted text.

Author 1: Monica Bhutani

Author 2: Mohammad Shuaib Mir

Author 3: Choo Wou Onn

Author 4: Yonis Gulzar

Keywords: Machine learning; data analysis; natural language processing; hierarchical transformer; context-aware computing; medical text mining; clinical decision support; healthcare; unstructured data processing

Paper 7: A Multi-Stage Detection of Diabetic Retinopathy in Fundus Images Using Convolutional Neural Network

Abstract: Diabetic Retinopathy (DRY) is a microvascular complication caused by diabetes mellitus, and it is one of the leading causes of blindness, especially in human adults. As the prevalence of this disease is growing exponentially, the screening of millions of people needs to be performed at a proliferating rate to diagnose the stage of the disease in its early stages. Highly advanced in the domain of technology, especially in artificial intelligence and its allied techniques, has come for the screening of DRY in photography to enhance the quality of life. This generates a bulk size of data that travels at high speed and cuts down on many human tasks. However, the techniques employed by the authors so far are quite expensive and time-consuming, and the prediction rate is insufficient to apply in a real-time scenario. This study offered a road for a deep learning-based fully automated system that helps to save manual disease diagnosis work and achieve disease detection in its very early stage using EfficientNetB3 (ENB3) Convolutional Neural Network (CNN) on DRY Fundus Images (FDI). In the suggested CNN, architectural variations and pre-processing techniques such as dimensionality reduction, global average pooling, and circular cropping are introduced alongside the Leaky ReLU (LR) activation function, Transfer Learning, and Reduce LROnPlateau technique, respectively. The accuracy of the proposed CNN classifier was 94.2% on training data, with a kappa score of 0.874, while it achieved a high level of accuracy at 96.7% on the testing data for DRY grading. Further, the evaluation results presented that the proposed model efficiently classifies the DRY stages for early disease detection.

Author 1: Puneet Kumar

Author 2: Salil Bharany

Author 3: Ateeq Ur Rehman

Author 4: Arjumand Bono Soomro

Author 5: Mohammad Shuaib Mir

Author 6: Yonis Gulzar

Keywords: Diabetic retinopathy; convolutional neural network; EfficientNetB3; fundus images; deep learning; transfer learning

Paper 8: Ontology-Based Automatic Generation of Learning Materials for Python Programming

Abstract: Learning materials in programming education are essential for effective instruction. This study introduces an ontology-based approach for automatically generating learning materials for Python programming. The method harnesses ontologies to capture domain knowledge and semantic relationships, enabling the creation of personalized, adaptive content. The ontology serves as a knowledge base to identify key concepts and resources and map them to learning objectives aligned with user preferences. The study outlines the design of a dual-module ontology: a general and a specific domain-specific concepts module. This design supports enhanced, tailored learning experiences, enhancing Python education by meeting individual needs and learning styles. The approach also increases the quality and uniformity of generated content, which can be reused for educational reasons. The system ensures alignment with reference materials by using BERT embeddings for a semantic similarity measurement, achieving a quality accuracy of 98.5%. It can be applied to improve Python education by providing personalized recommendations, hints, and problem-solution generation. Future developments could further support the functionality to strengthen teaching and learning outcomes in programming education, and it could expand to automated problem generation.

Author 1: Jawad Alshboul

Author 2: Erika Baksa-Varga

Keywords: Ontology; knowledge graph; learning material generation; domain knowledge; python

Paper 9: Artificial Intelligence Based System for Sorting and Detection of Organic and Inorganic Waste

Abstract: Solid waste management has become a global challenge today due to its constant increase in waste and inadequate classification, which leads to serious environmental problems. The research objective is to develop a system based on artificial intelligence (AI) for the classification and detection of organic and inorganic waste. In terms of its approach, it is quantitative with a pre-experimental and applied design. The population was made up of 1,298 images as a data collection technique for observation. Furthermore, the implementation of this system has shown significant improvements in its key indicators: precision, detection speed, and reduction of errors in the tests carried out, obtaining an increase in precision of 11.52%, 23.61% in detection speed and a reduction in 24.13% error rate. Finally, this research highlights the importance of AI in environmental sustainability by promoting much more efficient waste management and thus promoting ecological awareness in educational environments and for students to value the importance of recycling and sustainability. Finally, this research concludes that AI-based systems are a viable and scalable solution to address all the challenges associated with waste management.

Author 1: Angel Jair Castañeda Meza

Author 2: Nicol’s Alexander Lopez Haro

Author 3: Rosalynn Ornella Flores-Castañeda

Keywords: Artificial intelligence (AI); environmental sustainability; waste classification; organic waste; inorganic waste

Paper 10: Building Cyber-Resilient Universities: A Tailored Maturity Model for Strengthening Cybersecurity in Higher Education

Abstract: This study explores Higher Education Institutions (HEIs) cybersecurity maturity and preparedness, developing a Cybersecurity Maturity Model (CSMM) for HEIs specific to the needs of these institutions. These HEIs face increasing cyber threats and cyberattacks from ransomware attacks, phishing attempts, and data breaches, considering increasing dependence on digital methods for administration, teaching, and research. Though cybersecurity is of paramount importance today, many institutions do not have proper structures with which they can evaluate and enhance their security practices. The study uses a mixed-method approach, whereby the integration of qualitative case studies and quantitative surveys would address this gap, subsequently allowing the identification, validation, and assessment of the key domains and criteria in a comprehensive cybersecurity framework. The research started with an investigation, followed by design, data collection, analysis, and reporting, which accounted for the major phases of the study. The data was collected through interviews, documentation reviews, and surveys involving cybersecurity experts and ICT management teams in various HEIs. The results revealed eleven important assessment domains, twenty-four criteria, and sixty-seven elements necessary for developing the CSMM: Governance, Risk Management, Infrastructure Security, Human Factors, Compliance, and Monitoring. The validation confirmed the model to be practical, reliable, and valuable in the overall sense, giving the institutions a structured avenue for assessing and improving their cybersecurity maturity.

Author 1: Maznifah Salam

Author 2: Khairul Azmi Abu Bakar

Author 3: Azana Hafizah Mohd Aman

Keywords: Cybersecurity; HEIs; cybersecurity maturity model; mixed-method; governance

Paper 11: Automated Classification of Parasitic Worm Eggs Based on Transfer Learning and Fine-Tuned CNN Models

Abstract: Classification of worm eggs is important for diagnosing worm diseases, but the manual process is time-consuming. This study designs an image classification system using Convolutional Neural Network (CNN), transfer learning, and fine-tuning. The main goal of this study is to create a CNN model to sort parasitic worm eggs into groups. It does this by comparing three CNN architectures: EfficientNetB0, MobileNetV3, and ResNet50; it also creates classification technology for diagnosing worm infections. We applied transfer learning with pre-trained models and fine-tuned them for the IEEE parasitic egg dataset. The results reveal that EfficientNetB0 is superior, with an accuracy of 95.36%, precision of 95.80%, recall of 95.38%, and F1-score of 95.48%. It performs better and more efficiently than the other two architectures. Applying transfer learning and fine-tuning improves model performance, with EfficientNetB0 consistently outperforming. Furthermore, visual similarities between classes in the dataset likely cause prediction errors. Therefore, this system can support the diagnosis of worm diseases with high efficiency and accuracy.

Author 1: Ira Puspita Sari

Author 2: Budi Warsito

Author 3: Oky Dwi Nurhayati

Keywords: Classification; Convolutional Neural Network; EfficientNetB0; MobileNetV3; ResNet50

Paper 12: Evaluating Large Language Model Versus Human Performance in Islamophobia Dataset Annotation

Abstract: Manual annotation of large datasets is a time-consuming and resource-intensive process. Hiring annotators or outsourcing to specialized platforms can be costly, particularly for datasets requiring domain-specific expertise. Additionally, human annotation may introduce inconsistencies, especially when dealing with complex or ambiguous data, as interpretations can vary among annotators. Large Language Models (LLMs) offer a promising alternative by automating data annotation, potentially improving scalability and consistency. This study evaluates the performance of ChatGPT compared to human annotators in annotating an Islamophobia dataset. The dataset consists of fifty tweets from the X platform using the keywords Islam, Muslim, hijab, stopislam, jihadist, extremist, and terrorism. Human annotators, including experts in Islamic studies, linguistics, and clinical psychology, serve as a benchmark for accuracy. Cohen’s Kappa was used to measure agreement between LLM and human annotators. The results show substantial agreement between LLM and language experts (0.653) and clinical psychologists (0.638), while agreement with Islamic studies experts was fair (0.353). Overall, LLM demonstrated a substantial agreement (0.632) with all human annotators. ChatGPT achieved an overall accuracy of 82%, a recall of 69.5%, an F1-score of 77.2%, and a precision of 88%, indicating strong effectiveness in identifying Islamophobia-related content. The findings suggest that LLMs can effectively detect Islamophobic content and serve as valuable tools for preliminary screenings or as complementary aids to human annotation. Through this analysis, the study seeks to understand the strengths and limitations of LLMs in handling nuanced and culturally sensitive data, contributing to broader discussion on the integration of generative AI in annotation tasks. While LLMs show great potential in sentiment analysis, challenges remain in interpreting context-specific nuances. This study underscores the role of generative AI in enhancing human annotation efforts while highlighting the need for continuous improvements to optimize performance.

Author 1: Rafizah Daud

Author 2: Nurlida Basir

Author 3: Nur Fatin Nabila Mohd Rafei Heng

Author 4: Meor Mohd Shahrulnizam Meor Sepli

Author 5: Melinda Melinda

Keywords: Large Language Model; generative AI; human intelligence; automatic data annotation; sentiment analysis; islamophobia; ChatGPT

Paper 13: Exploring the Landscape of 6G Wireless Communication Technology: A Review

Abstract: The advent of 6G technology promises to revolutionize the landscape of connectivity, ushering in an era of unprecedented speed, reliability, and integration of emerging technologies. This comprehensive review delves into the evolving domain of 6G wireless communication technology, synthesizing current research, trends, and projections to provide a holistic understanding of its potential impact and challenges. Beginning with an overview of the evolution from previous generations, the review examines the foundational principles, key features, and technological advancements envisioned for 6G networks. It explores concepts such as terahertz communication, ultra-reliable low latency communication (URLLC), intelligent surfaces, and holographic beamforming, elucidating their potential to redefine communication paradigms. The integration of artificial intelligence (AI) and edge computing is highlighted as pivotal in enabling intelligent, adaptive, and efficient network operations. Furthermore, the review investigates how 6G is expected to support massive-scale Internet of Things (IoT) deployments and considers the future role of quantum computing in enhancing security and processing capabilities. Regulatory and standardization frameworks essential for the development and deployment of 6G networks are scrutinized, alongside addressing issues concerning security, privacy, and sustainability. By synthesizing insights from academia, industry, and standardization bodies, this review provides a roadmap for researchers, policymakers, and industry stakeholders to navigate the evolving landscape of 6G and realize its transformative potential in shaping the future of global connectivity.

Author 1: Nur Arzilawati Md Yunus

Author 2: Zurina Mohd Hanapi

Author 3: Shafinah Kamarudin

Author 4: Aindurar Rania Balqis Mohd Sufian

Author 5: Fazlina Mohd Ali

Author 6: Nabilah Ripin

Author 7: Hazrina Sofian

Keywords: 6G; wireless communication technology; artificial intelligence; connectivity; edge computing; Internet of Things (IoT); quantum computing; terahertz communication; Ultra-Reliable Low Latency Communication (URLLC)

Paper 14: Human Detection and Tracking with YOLO and SORT Tracking Algorithm

Abstract: Human tracking is often performed on publicly available well annotated datasets, where the dataset development is always avoided because of the tiring process. Publicly available well-annotated datasets are ideal for training because those generate higher tracking accuracy. This study performs human tracking on videos recorded manually using optimized detectors following the tracking by detection framework. Manually recorded videos were used to develop a dataset which comprises more than 8k image sequences. Both indoor and outdoor scenarios were chosen to maintain different lighting conditions which make tracking difficult. All these image frames are labelled with bounding boxes for humans. The dataset is prepared by following the MOT15 dataset structure. A unique annotation process was performed that reduced the annotation labor by almost 80% which was a combination of manual annotation and prediction from pretrained models. Different sizes of You Only Look Once (YOLO) detection model (n/s/m) were trained using the train dataset focusing on humans and coupled with two most popular tracking algorithms: Simple Online Realtime Tracking (SORT) and DeepSORT. The YOLOv8 and YOLO11 models were optimized with proper hyperparameter values followed by tracking, using SORT and DeepSORT. The results were observed with those models on different confidence and Intersection over Union (IoU) threshold values. This study finds a proportional relation with the optimization of detection models and tracking accuracy. YOLO11m with DeepSORT tracker performed best on the test data with 74% Multiple Object Tracking Accuracy (MOTA) also the other optimized YOLO models tend to perform better with the trackers than the unoptimized ones.

Author 1: Tanveer Kader

Author 2: Ahmad Fakhri Ab. Nasir

Author 3: M. Zulfahmi Toh

Author 4: Muhammad Nur Aiman Shapiee

Author 5: Amir Fakarullsroq Abdul Razak

Keywords: Human tracking; multiple object tracking; tracking-by-detection; you only look once (YOLO); simple online and realtime tracking (SORT)

Paper 15: Optimal Algorithm of Expressway Maintenance Scheme Based on Genetic Algorithm

Abstract: The genetic algorithm (GA), characterized by parallelism and global optimization capabilities, is well-suited for solving optimization problems related to expressway maintenance schemes. In this study, we improved GA operators and algorithm parameters within the existing maintenance scheme optimization model, thereby enhancing the operational efficiency of the GA. Building on this foundation, an optimization algorithm for expressway maintenance schemes was developed. Subsequently, MATLAB was employed to program the algorithm and solve the expressway maintenance scheme problem. When compared with the solution results in the reference, the proposed approach achieved a reduction of approximately 3.6% in maintenance costs and an improvement of about 47% in operation speed, verifying the algorithm's reliability and effectiveness. Finally, visualization of the algorithm program was enabled using MATLAB App Designer and MATLAB Compiler. This method can be popularized and applied in aspects such as expressway maintenance decision-making and optimization of building maintenance schemes.

Author 1: Yushu Zhu

Author 2: Xingwang Liu

Author 3: Fengshuang Zhang

Author 4: Kashan Khan

Author 5: Yang Chen

Author 6: Runqi Liu

Author 7: Qiang He

Keywords: Genetic algorithm (GA); expressway; scheme optimization; MATLAB; program development

Paper 16: Pet Cat Home Design Evaluation System: Based On Grounded Theory-CRITIC-TOPSIS

Abstract: As pet cats assume an increasingly significant role in households, the variety of pet-cat home products on the market has proliferated. However, existing studies primarily focus on qualitative assessments of individual product functions or user experiences, and lack a systematic evaluation framework that combines in-depth exploration of user needs with quantitative analysis. To address this research gap and with the objectives of enhancing user satisfaction and guiding product development, this study constructs a user-needs–based evaluation framework for pet-cat home design. Semi-structured interviews with 12 pet-cat owners were conducted and analyzed via Grounded Theory to elicit four core requirements—Enhancing Pet Life Quality (A1), Ease of Cleaning and Maintenance (A2), Aesthetic Appeal (A3), and Safety and Reliability (A4)—and thirteen primary requirement elements. The CRITIC method was then applied to determine the weights of these dimensions (A1 = 0.30, A2 = 0.28, A3 = 0.27, A4 = 0.16). Four representative market products were selected and ranked using the TOPSIS method based on their proximity to the ideal and negative-ideal solutions, quantitatively evaluating their relative merits. Results indicate that pet owners prioritize Enhancing Pet Life Quality and Ease of Cleaning and Maintenance (combined weight = 0.58), providing focused guidance for designers on spatial layout and material selection. Aesthetic Appeal and Safety and Reliability also remain critical, pointing to specific optimization directions for product appearance and structural integrity. This study not only fills a methodological gap in pet-cat home design evaluation but also offers a practical model for weighting user needs and selecting optimal design solutions, thereby contributing to the standardization and refinement of pet home products.

Author 1: Yuzhe Qi

Author 2: Hengwang Zhang

Author 3: Yaping Liu

Keywords: Grounded theory; CRITIC; TOPSIS; design evaluation; pet home

Paper 17: Enhanced Bidirectional LSTM for Sentiment Analysis of Learners’ Posts in MOOCs

Abstract: Massive Open Online Courses (MOOCs) have transformed digital learning, leading to vast amounts of learner-generated content that reflect user experience and engagement. Accurately classifying sentiment from this content is essential for improving course quality, but remains challenging due to subtle linguistic variation and contextual ambiguity. This study proposes a sentiment analysis approach based on an enhanced Bidirectional Long Short-Term Memory (LSTM) model. The enhancements include the integration of data augmentation and regularization techniques to address overfitting and improve generalization. The model was trained and evaluated on a dataset of 29,604 learner discussion posts from Stanford University MOOCs. Experimental results show that the proposed model achieves an accuracy of 88.54% in classifying sentiments into positive, negative, and neutral classes. These results suggest that the enhanced LSTM model offers a reliable solution for large-scale sentiment classification in online education, with potential applications in learner support, curriculum design, and personalized feedback.

Author 1: Chakir Fri

Author 2: Rachid Elouahbi

Author 3: Youssef Taki

Author 4: Ahmed Remaida

Keywords: MOOCs; Sentiment analysis; deep learning; Bidirectional LSTM; data augmentation; regularization techniques

Paper 18: Exploring Research Trends in Distributed Acoustic Sensing with Machine Learning and Deep Learning: A Bibliometric Analysis of Themes and Emerging Topics

Abstract: This paper explores the emerging research trends in Distributed Acoustic Sensing (DAS) with the integration of Machine Learning and Deep Learning technologies. DAS has diverse applications, including subsurface seismic monitoring, pipeline surveillance, and natural disaster detection. Using the Scopus database, 323 documents published between 2011 and 2023 were analysed. Through a comprehensive bibliometric analysis using the “bibliometrix” R package, the study aims to document the advancement in DAS techniques over the last decade, highlighting the publication patterns, key contributors, and frequently explored themes. The analysis reveals a steady increase in research output, with significant contributions from China and the United States. Core research areas identified include seismic monitoring, pipeline security, and infrastructure health monitoring. Additionally, the paper examines the impact of key publications, influential authors, and prolific research institutions. The findings provide valuable insights for both academic and industrial stakeholders, underscoring the potential for future innovations in DAS applications and helping to identify potential research gaps.

Author 1: Nor Farisha Muhamad Krishnan

Author 2: Jafreezal Jaafar

Keywords: Machine learning; deep learning; distributed acoustic sensing; bibliometric

Paper 19: Nonlinear Consensus for Wireless Sensor Networks: Enhancing Convergence in Neighbor-Influenced Models

Abstract: Wireless sensor networks (WSNs) are a modern technology that has revolutionized many industries thanks to their ability to collect and analyze information from surrounding environments and improve the performance of complex systems through the cooperation of a group of independent sensors to achieve common goals. Sensor clustering and agreement have wide applications in daily life, ranging from environmental monitoring and industrial control to healthcare and smart cities. However, the WSN system faces many challenges, one of the most prominent is achieving agreement between different sensors on a common state. This challenge is essential to enable successful cooperation between sensors in complex systems. Many previous research and models have been developed to address the problem of sensor agreement, such as the Neighbor-Influenced Timestep Consensus Model (NITCM), which was presented as a framework to achieve agreement effectively. In this paper, we propose a new technique to improve this model by using fractional force in the updating process. This leads to developing the Neighbor-Influenced Fractional Timestep Consensus Model (NIFTCM). This technique achieves faster convergence between sensors, which leads to improved efficiency in reaching agreement over previous techniques. This development aims to enhance the speed and stability of consensus processes in wireless sensor networks and make them more suitable for time-sensitive applications.

Author 1: Rawad Abdulghafor

Author 2: Yousuf Al Husaini

Author 3: Abdullah Said AL-Aamri

Author 4: Mohammad Abrar

Author 5: Alaa A. K. Ismaeel

Author 6: Mohammed Abdulla Salim Al Husaini

Keywords: Fractional power; consensus; WSNs; NIAM; NIFFAM

Paper 20: Breast Cancer Classification and Segmentation Using Deep Learning on Ultrasound Images

Abstract: Breast cancer continues to pose a major health challenge for women worldwide, highlighting the critical role of accurate and early detection methods in improving patient outcomes. Ultrasound imaging, a commonly used and non-invasive method, is especially useful for identifying tissue irregularities in younger women or individuals with dense breast tissue. However, accurate interpretation of ultrasound images is challenging due to variability in human analysis and limitations in existing deep learning models, which often struggle with small, imbalanced datasets and lack generalizability compared to models trained on natural images. To tackle these challenges, we introduce a dual deep learning framework that combines image classification and tumor segmentation using breast ultrasound images. The classification component evaluates four models (Custom CNN, VGG16, InceptionV3, and MobileNet) while the segmentation module employs a MobileNet-optimized U-Net architecture for precise boundary localization. We validate our approach using the publicly available BUSI dataset, achieving a 98% classification accuracy with MobileNet and a Dice coefficient of 0.8959 for segmentation, indicating high model reliability and spatial agreement. Our method demonstrates a robust, efficient solution to automate breast cancer detection and localization, with potential to support radiologists in early and accurate diagnosis.

Author 1: Doha Saad Dajam

Author 2: Ayman Qahmash

Keywords: Breast cancer; Convolutional Neural Networks (CNNs); tumor segmentation; MobileNet; dice coefficient; BUSI Dataset

Paper 21: DamageNet: A Dilated Convolution Feature Pyramid Network Mask R‑CNN for Automated Car Damage Detection and Segmentation

Abstract: Automated and precise assessment of vehicle damage is critical for modern insurance processing, accident analysis, and autonomous maintenance systems. In this work, we introduce DamageNet, a unified deep instance segmentation framework that embeds a multi‑rate dilated‑convolution context module within a Feature Pyramid Network (FPN) backbone and couples it with a Region Proposal Network (RPN), RoI‑Align, and parallel heads for classification, bounding‑box regression, and pixel‑level mask prediction. Evaluated on the large‑scale VehiDE dataset comprising 5 200 high‑resolution images annotated for dents, scratches, and broken glass, DamageNet achieves a mean Average Precision (mAP) of 85.7% for damage localization and a mean Intersection over Union (mIoU) of 82.3% for segmentation, outperforming baseline Mask R‑CNN by 6.2 and 7.8 percentage points, respectively. Ablation studies confirm that the dilated‑convolution module, multi‑scale fusion in the FPN, and post‑processing refinements each contribute substantially to segmentation fidelity. Qualitative results demonstrate robust delineation of both subtle scratch lines and extensive panel deformations under diverse lighting and occlusion conditions. Although the integration of atrous convolutions introduces a modest inference overhead, DamageNet offers a significant advancement in end‑to‑end vehicle damage analysis. Future extensions will investigate lightweight dilation approximations, dynamic rate selection, and semi‑supervised learning strategies to further enhance processing speed and generalization to additional damage modalities.

Author 1: Nazbek Katayev

Author 2: Zhanna Yessengaliyeva

Author 3: Zhazira Kozhamkulova

Author 4: Zhanel Bakirova

Author 5: Assylzat Abuova

Author 6: Gulbagila Kuandikova

Keywords: Car damage detection; instance segmentation; dilated convolution; feature pyramid network; Mask R‑CNN; deep learning; vehicle damage assessment; semantic segmentation

Paper 22: Hybrid Structure Query Language Injection (SQLi) Detection Using Deep Q-Networks: A Reinforcement Machine Learning Model

Abstract: Structured Query Language injection (SQLi) remains one of the most pervasive and dangerous threats to web-based systems, capable of compromising databases and bypassing authentication protocols. Despite advancements in machine learning for cybersecurity, many models rely on static detection rules or require extensive labeled datasets, making them less adaptable to evolving threats. Addressing this limitation, the present study aimed to design, implement, and evaluate a Deep Q-Network (DQN) model capable of detecting SQLi attacks using reinforcement learning. The research employed a Design and Development Research (DDR) methodology, supported by an evolutionary prototyping framework, and utilized a dataset of 30,919 labeled SQL queries, balanced between malicious and safe inputs. Preprocessing involved query normalization and vector encoding into fixed-length ASCII representations. The DQN model was trained over 2,000 episodes, using experience replay and an epsilon-greedy strategy. Key evaluation metrics—accuracy, cumulative reward, and epsilon decay—showed performance improvements, with accuracy increasing from 52% to 82% and stabilizing between 65% and 73% in later episodes. The agent demonstrated consistent adaptability by successfully generalizing across various injection patterns. This outcome suggests that reinforcement learning, particularly using DQN, provides a viable alternative to traditional models, with superior resilience and dynamic learning capabilities. The model's convergence trend highlights its practical application in real-time SQLi detection systems, contributing significantly to cybersecurity measures for database-driven applications.

Author 1: Carlo Jude P. Abuda

Author 2: Cristina E. Dumdumaya

Keywords: Adaptive systems; cybersecurity; deep q-network; intrusion detection; query classification; reinforcement learning; SQL injection

Paper 23: Robot Path Planning Model Based on Improved A* Algorithm

Abstract: Robot path planning is a key technology for achieving autonomous navigation and efficient operation of robots. In order to improve the autonomous navigation capability of mobile robots, a global path planning model based on an improved A* algorithm and a local path planning model based on an improved artificial potential field method were designed. The results showed that the turns in the optimal path under the improved A* algorithm were 8, 5, 9, and 5, respectively. The improved artificial potential field method achieved a maximum planning time of 0.17s and a minimum planning time of 0.11s. The designed global and local path planning models for mobile robots have good performance and can provide technical support for improving the autonomous navigation capability of mobile robots for industrial manufacturing.

Author 1: Jing Xie

Author 2: Chunyuan Xu

Author 3: Qianxi Yang

Keywords: Robot; path; planning; A* algorithm; artificial potential field method; SA

Paper 24: Integrating ISA Optimised Random Forest Methods for Building Applications in Digital Accounting Talent Assessment

Abstract: Digital accounting talent assessment in applied undergraduate colleges and universities is an urgent problem of talent assessment construction. In order to solve the problem of digital accounting talent assessment in applied undergraduate colleges, a digital accounting talent assessment method based on improved machine learning algorithm is proposed. Firstly, the digital accounting talent assessment problem in applied undergraduate colleges is analysed, digital accounting talent assessment indicators are extracted, and the index system is constructed; secondly, the digital accounting talent assessment model based on the integrated ISA optimized random forest algorithm in applied undergraduate colleges is constructed by combining the integrated learning technology, the intelligent optimization algorithm, and the random forest; lastly, the digital accounting talent data in applied undergraduate colleges is used to analyse the model. The results show that compared with other algorithms, the accuracy of digital accounting talent assessment in applied undergraduate colleges and universities of Ada-ISA-RF is improved by 3.06 per cent and 7.04 per cent, respectively.

Author 1: Yu ZHOU

Keywords: Integrated learning; internal renovation algorithms; random forests; digitalisation of applied undergraduate institutions; accounting talent assessment

Paper 25: Binary–Source Code Matching Based on Decompilation Techniques and Graph Analysis

Abstract: Recent approaches to binary–source code matching often operate at the intermediate representation (IR) level, with some applying the matching process at the binary level by compiling the source code to binary and then matching it directly with the binary code. Others, though less common, perform matching at the decompiler-generated pseudo-code level by first decompiling the binary code into pseudo-code and then comparing it with the source code. However, all these approaches are limited by the loss of semantic information in the original source code and the introduction of noise during compilation and decompilation, making accurate matching challenging and often requiring specialized expertise. To address these limitations, this study introduces a system for binary–source code matching based on decompilation techniques and Graph analysis (BSMDG) that matches binary code with source code at the source code level. Our method utilizes the Ghidra decompiler in conjunction with a custom-built transpiler to reconstruct high-level C++ source code from binary executables. Subsequently, call graphs (CGs) and control flow graphs (CFGs) are generated for both the original and translated code to evaluate their structural and semantic similarities. To evaluate our system, we used a curated dataset of C++ source code and corresponding binary files collected from the AtCoder website for training and testing. Additionally, a case study was conducted using the widely recognized POJ-104 benchmark dataset to assess the system's generalizability. The results demonstrate the effectiveness of combining decompilation with graph-based analysis, with our system achieving 90% accuracy on POJ-104, highlighting its potential in code clone detection, vulnerability identification, and reverse engineering tasks.

Author 1: Ghader Aljebreen

Author 2: Reem Alnanih

Author 3: Fathy Eassa

Author 4: Maher Khemakhem

Author 5: Kamal Jambi

Author 6: Muhammed Usman Ashraf

Keywords: Binary–source code matching; call graphs; code clone detection; control flow graphs; decompiler

Paper 26: Computational Linguistic Approach for Holistic User Behaviors Modeling Through Opinionated Data of Virtual Communities

Abstract: This research is aimed at establishing a computational linguistic model for the detection of positive and negative statements, synthesized for the Pakistani microblogging site Twitter, particularly, in the Roman Urdu language. With increased freedom of speech people express their sentiments towards an event or a person in positive, negative, neutral, and sometimes sarcastic tones, especially on social media platforms. Pakistani social media users, like other multilingual countries, express their opinions through code switching and code mixing. Their language lacks correct grammar, informal and nonstandard writing, unrelated spelling, alternative analogies make it difficult for computational linguist to mine their data for computational research. To overcome this challenge, the study employed web scraping tools to retrieve a large number of Roman Urdu tweets. In order to establish a new positive and negative statements corpus, the text data is annotated through a sentiment analysis carried out by using TextBlob sentiment analysis and Bidirectional Encoder Representations from Transformers (BERT). Addressing this issue makes it possible to eliminate the gap that is evident in the models that do not identify Roman Urdu as a form of language. The findings are useful for the regulatory bodies and researchers since it offers a culturally and linguistically appropriate database and model targeting resource constraints and key performance metrics. It helps in content moderation and in making policies regarding the technological advancement within Pakistan.

Author 1: Kashif Asrar

Author 2: Syed Abbas Ali

Keywords: Roman Urdu; positive and negative statements detection; sentiment analysis; BERT Model; Long Short-Term Memory (LSTM) networks; Pakistani social media

Paper 27: Securing UAV Flight Data Using Lightweight Cryptography and Image Steganography

Abstract: The popularity of Unmanned Aerial Vehicles (UAVs) in various fields has been rising recently. UAV technology is being invested in by numerous industries in order to cut expenses and increase efficiency. Therefore, UAVs are predicted to become much more important in the future. As UAVs become more popular, the risk of cyberattacks on them is also growing. One type of cyberattack involves the exposure of important flight data. This, in turn, can lead to serious problems. To address this problem, a new method based on lightweight cryptography and steganography is proposed in this work. The proposed method ensures multilayer protection of important UAV flight data. This is achieved by two layers of encryption using a polyalphabetic substitution cipher and ChaCha20-Poly1305 authenticated encryption, as well as randomized least significant bit (LSB) steganography. Most importantly, through this work, a balance is kept between security and performance. Additionally, all experiments are carried out on real devices, making the proposed method more practical. The proposed method is evaluated using MSE, PSNR, and SSIM metrics. Even with a capacity of 8000 bytes, it achieves an MSE of 0.04, a PSNR of 62, and an SSIM of 0.9998. It is then compared to existing methods. The results show better practical use, stronger security, and higher overall performance.

Author 1: Orkhan Valikhanli

Author 2: Fargana Abdullayeva

Keywords: UAV; GCS; cyberattack; cryptography; steganography; flight data

Paper 28: Instance Segmentation Method Based on DPA-SOLOV2

Abstract: To solve the problems of missed detection, segmentation errors in instance segmentation models, we propose an instance segmentation approach, DPA-SOLOV2, based on the improved segmenting objects by locations V2 (SOLO V2). Firstly, DPA-SOLOV2 introduces deformable convolutional networks (DCN) into the feature extraction network ResNet50. By freely sampling points to convolve features of any shape, the network can extract feature information more effectively. Secondly, DPA-SOLOV2 uses the path aggregation feature pyramid network (PAFPN) feature fusion method to replace the feature pyramid. By adding a bottom-up path, it can better transmit the location information of features and also enhance the information interaction between features. To prove the effectiveness of the improved model, we conduct experiments on two public datasets, COCO and CVPPP. The experimental results show that the accuracy of the improved model on the COCO dataset is 1.3% higher than that of the original model, and the accuracy on the CVPPP dataset is 1.5% higher than that before the improvement. Finally, the improved model is applied to the insulator dataset, which can accurately segment the umbrella skirt of insulators and outperforms other mainstream instance segmentation algorithms such as Yolact++.

Author 1: Yuyue Feng

Author 2: Liqun Ma

Author 3: Yinbao Xie

Author 4: Zhijian Qu

Keywords: Instance segmentation; segmenting objects by locations V2; deformable convolutional networks; path aggregation feature pyramid network; insulator dataset

Paper 29: Reducing Cyber Violence and Fostering Empathy Through VRN4RCV Model: Expert Review

Abstract: Cyber violence has become increasingly prevalent, necessitating innovative intervention strategies. VR technology, with its immersive and empathetic capabilities, provides a unique opportunity for influencing behavioral change among perpetrators of cyber violence. This study proposes a conceptual design model for VR news, aimed at fostering empathy through immersive experiences to reduce cyber violence. The model was validated through three cycles of expert review. Expert feedback highlighted the model’s relevance and applicability while offering constructive suggestions for refinement. The findings indicate that this conceptual model provides a practical guide for designing VR news that effectively addresses the issue of cyber violence. Future research will include prototype testing and empirical evaluation to assess the model’s impact on behavioral change and empathy enhancement.

Author 1: Wu Qiong

Author 2: Nadia Diyana Binti Mohd Muhaiyuddin

Author 3: Azliza Binti Othman

Keywords: Cyber violence; VR news; empathy; conceptual model; expert review

Paper 30: Systematic Literature Review on Generative AI: Ethical Challenges and Opportunities

Abstract: Generative Artificial Intelligence (GAI) has rapidly emerged as a transformative technology capable of autonomously creating human-like content across domains such as text, images, code, and media. While GAI offers significant benefits in fields like education, healthcare, and creative industries, it also introduces complex ethical challenges. This study aims to systematically review and synthesize the ethical landscape of GAI by analyzing 112 peer-reviewed journal articles published between 2021 and 2025. Using a Systematic Literature Review (SLR) methodology, the study identifies five primary ethical challenges—bias and discrimination, misinformation and deepfakes, data privacy violations, intellectual property issues, and accountability and explainability. In addition, it highlights emerging opportunities for ethical innovation, such as responsible design, inclusive governance, and interdisciplinary collaboration. The findings reveal a fragmented research landscape with limited empirical validation and inconsistent ethical frameworks. This review contributes to the field by mapping cross-sectoral patterns, identifying critical research gaps, and offering practical directions for researchers, developers, and policymakers to promote the responsible development of generative AI.

Author 1: Feliks Prasepta Sejahtera Surbakti

Keywords: Generative Artificial Intelligence (GAI); AI ethics; systematic literature review; bias; misinformation; data privacy; accountability

Paper 31: A Deep Learning Model for Speech Emotion Recognition on RAVDESS Dataset

Abstract: Speech Emotion Recognition (SER), a pivotal area in artificial intelligence, is dedicated to analyzing and interpreting emotional information in human speech. To address the challenges of capturing both local acoustic features and long-range dependencies in emotional speech, this study proposes a novel parallel neural network architecture that integrates Convolutional Neural Networks (CNNs) and Transformer encoders. To integrate the distinct feature representations captured by the two branches, a cross-attention mechanism is employed for feature-level fusion, enabling deep-level semantic interaction and enhancing the model’s emotion discrimination capacity. To improve model generalization and robustness, a systematic preprocessing pipeline is constructed, including signal normalization, data segmentation, additive white Gaussian noise (AWGN) augmentation with varying SNR levels, and Mel spectrogram feature extraction. A grid search strategy is adopted to optimize key hyperparameters such as learning rate, dropout rate, and batch size. Extensive experiments conducted on the RAVDESS dataset, consisting of eight emotional categories, demonstrate that our model achieves an overall accuracy of 80.00%, surpassing existing methods such as CNN-based (71.61%), multilingual CNN (77.60%), bimodal LSTM-attention (65.42%), and unsupervised feature learning (69.06%) models. Further analyses reveal its robustness across different gender groups and emotional intensities. Such outcomes highlight the architectural soundness of our model and underscore its potential to inform subsequent developments in affective speech processing.

Author 1: Zhongliang Wei

Author 2: Chang Ge

Author 3: Chang Su

Author 4: Ruofan Chen

Author 5: Jing Sun

Keywords: Speech emotion recognition; deep learning; RAVDESS dataset; multi-feature fusion

Paper 32: Design and Evaluation of a Forensic-Ready Framework for Smart Classrooms

Abstract: The rise of cyber threats in educational environments underscores the need for forensic-ready systems tailored to digital learning platforms like smart classrooms. This study proposes a proactive forensic-ready framework that integrates threat estimation, risk profiling, data identification, and collection management into a continuous readiness cycle. Blockchain technology ensures log immutability, while LMS APIs enable systematic evidence capture with minimal disruption to learning processes. Monte Carlo Simulation validates the framework’s performance across key metrics. Results show a log capture success rate of 77.27%, with high accuracy for structured attacks such as SQL Injection. The system maintains operational efficiency, adding only 15% average CPU overhead. Forensic logs are securely stored in JSON format on a blockchain ledger, ensuring both integrity and accessibility. However, reduced effectiveness for complex attacks like Remote Code Execution and occasional retrieval delays under heavy loads highlight areas for improvement. Future enhancements will focus on expanding threat coverage and optimizing log retrieval. By addressing vulnerabilities unique to smart classrooms, such as unauthorized access and data manipulation, this study introduces a scalable, domain-specific solution for enhancing forensic readiness and cybersecurity in educational ecosystems.

Author 1: Henry Rossi Andrian

Author 2: Suhardi

Author 3: I Gusti Bagus Baskara Nugraha

Keywords: Forensic-ready system; smart classroom; threat estimation; risk profile

Paper 33: Method for Effect Evaluation of a Reception System on Sales, Number of Customers, Hourly Productivity and Churn Based on Intervention Analysis

Abstract: We propose a method of AI-based evaluation of sales, number of customers, and churn before and after the introduction of a hair salon based on intervention time series analysis. We also used the software package of CausalImpact for the intervention time series analysis. The problem with this method is that the prediction accuracy is insufficient, and the estimated results of the intervention effect are not very valid. We thought it was necessary to verify prediction accuracy by using data before the system was introduced, where correct answer data exists, for the counterfactual prediction value after the system was introduced and devised a method to accurately predict the outcome variable before the system was introduced. Specifically, we introduce two learning models as in the development workflow of a general machine learning model, one for learning and the other one for accuracy verification. However, since CausalImpact does not include the function to verify the prediction accuracy, a separate code was prepared for that purpose to improve the prediction accuracy. As a result, we were able to confirm that the prediction accuracy was almost acceptable.

Author 1: Kohei Arai

Author 2: Ikuya Fujikawa

Author 3: Sayuri Ogawa

Keywords: Intervention time series analysis; causalimpact package; counterfactual prediction value; general machine learning model

Paper 34: Modified MobileNet-V2 Convolution Neural Network (CNN) for Character Identification of Surakarta Shadow Puppets

Abstract: Shadow puppets or in Indonesian called as “wayang kulit” is one of Indonesia's native traditional arts that still exists to this day. This art form has been recognised by UNESCO since 2003. Wayang kulit is not just ordinary entertainment. It carries profound moral values, but is gradually being forgotten by the younger generation. To facilitate the public in recognizing wayang kulit characters, a desktop-based application was developed using Canny edge detection for image extraction and a modified MobileNet-V2 CNN algorithm for character identification. The dataset used in this research was sourced from Google and Instagram, with 22 names of wayang kulit characters serving as classes. The identification results for 1,312 wayang kulit images (test data) using the classic CNN model yielded an accuracy of 50%, precision of 53%, and recall of 47%. Meanwhile, with the modified MobileNet-V2 CNN model, called custom CNN gives an accuracy of 92%, precision of 93%, and recall of 92%. From the result, it is shown that the custom CNN has high performance, where it has a few false positive predictions in detecting the characters of wayang kulit. Additionally, the result shows that the CNN model is robust and reliable for the task of identifying the wayang kulit characters. Based on the result, the model can be applied in preserving and promoting traditional wayang kulit art by helping to catalog and identify characters, making it more accessible to a wider audience, including the younger generation.

Author 1: Achmad Solichin

Author 2: Dwi Pebrianti

Author 3: Painem

Author 4: Sanding Riyanto

Keywords: Wayang kulit; characters identification; Convolution Neural Network (CNN); machine learning; image processing

Paper 35: Early Detection and Forecasting of Influenza Epidemics Using a Hybrid ARIMA-GRU Model

Abstract: Early diagnosis and accurate epidemic prediction are essential in limiting the public health impact of influenza epidemics because intervention on time can effectively curb both the spread of the disease and the strain on health services. Standard ARIMA models have proven their usefulness in short-term forecasting, particularly in stable contexts, but the fact that they cannot keep up with the complex and non-linear dynamics of disease spread makes them less capable of dealing with rapid-evolving outbreaks. This is especially the case when outbreaks are characterized by complicated seasonal trends and irregular peaks which are challenging for ARIMA to predict by itself. To fill this deficit, this study presents a hybrid model that marries ARIMA's statistical strength in dealing with short-term trends and the high-powered deep learning strengths of Gated Recurrent Units (GRU) that specialize in detecting long-term dependencies and non-linear relationships in data. The WHO Flu Net dataset, a trusted source of influenza surveillance, forms the foundation of training the model, with careful preprocessing operations conducted to normalize the data and eliminate any missing values, providing high-quality input to the model to make precise predictions. By combining ARIMA's linear prediction strengths with GRU's sophisticated pattern detection, the hybrid model delivers a powerful solution that is better than both regular ARIMA and other machine learning models, as evidenced by lower error rates on test metrics like MAE, RMSE and MAPE. The experimental findings validate that the ARIMA-GRU model not only enhances predictive performance but also increases the model's sensitivity to subtle trends, making it a valuable asset for early detection systems in public health. In the future, the incorporation of real-time environmental information such as temperature, humidity, and mobility patterns may further enhance the model's accuracy and responsiveness, providing more robust forecasting. Also, integrating healthcare infrastructure-related data, i.e., hospital capacity and availability of medical resources, would aid in developing a more complete epidemic management process. In total, the ARIMA-GRU hybridization is an effective and novel strategy for enhancing influenza surveillance, outbreak detection at the early stage, and epidemic control operations.

Author 1: Kabilan Annadurai

Author 2: Aanandha Saravanan

Author 3: S. Kayalvili

Author 4: Madhura K

Author 5: Elangovan Muniyandy

Author 6: Inakollu Aswani

Author 7: Yousef A.Baker El-Ebiary

Keywords: Time-series analysis; gated recurrent unit; temporal patterns; influenza epidemic; auto regressive integrated moving average; early detection

Paper 36: Survival Analysis and Machine Learning Models for Predicting Heart Failure Outcomes

Abstract: Heart failure is still one of the prominent causes of morbidity and mortality globally, and thus, determining the principal factors influencing survival in patients becomes crucial. Being able to predict survival is critical for optimizing patient treatment and management. Heart failure, with its multifactorial and involvement of numerous clinical variables, complicates prediction of survival rates in patients. This study utilizes the "Heart Failure Clinical Records" dataset to analyze and predict patient survival based on two separate approaches: survival analysis and machine learning (ML) classification. Specifically, we employ the Cox Proportional Hazards Model to assess the influence of clinical variables like “age”, “serum creatinine”, and “ejection fraction” on survival durations. Additionally, machine learning classification models like K-Nearest Neighbors (KNN), Decision Trees (DT), and Random Forests (RF) are implemented to predict the binary response variable of survival (DEATH_EVENT). Data preprocessing is carried out using methods like feature scaling, imputation of missing values, and balancing the classes for the improvement of model performance. Among the evaluated models, the Random Forest classifier, when integrated with feature selection derived from the Cox model, reached the best performance with 96.2% accuracy and an AUC ROC of 0.987, outperforming all other approaches. The results indicate that integrating survival analysis with machine-learning techniques is effective in heart failure prediction outcomes, providing valuable support for patient management and clinical decision-making.

Author 1: Naseem Mohammed ALQahtani

Author 2: Abdulmohsen Algarni

Keywords: Heart failure prediction; machine learning; cox proportional hazards model; random forest

Paper 37: Topology Planning and Optimization of DC Distribution Network Based on Mixed Integer Programming and Genetic Algorithm

Abstract: In the current situation of rapid development of the power industry, DC distribution network topology planning and optimization are of vital importance. This research studies the shortcomings of existing methods in terms of computational efficiency and optimization effect. Based on the real data of a medium-sized DC distribution network in a large city with 200 nodes and 350 lines, an innovative method combining mixed integer programming (MIP) and genetic algorithm (GA) is adopted. MIP is used to accurately describe physical constraints and optimization objectives, and GA efficiently searches for the best solution in the solution space with its global search capability. Experimental results show that the MIP-GA model has the lowest power transmission loss at different load levels. For example, at high load, it is 32% lower than the baseline, 16% lower than the MIP model, and 12.5% lower than the ACO model. It also performs best in terms of node voltage deviation, reliability, power quality and other indicators. Cost-benefit analysis shows that although the MIP-GA model has a relatively high investment cost for topology adjustment, it has the lowest annual power loss and maintenance cost, a reasonable total annual cost, a benefit-cost ratio of 1.5, and a payback period of only 3 years. Research has shown that this hybrid model has significant advantages in DC distribution network topology planning and optimization, and can effectively improve system performance and economic benefits.

Author 1: Ran Cheng

Author 2: Chong Gao

Author 3: Hao Li

Author 4: Junxiao Zhang

Author 5: Ye Huang

Keywords: DC distribution network; topology planning; mixed integer programming; genetic algorithm; optimization effect

Paper 38: Enhancing Customer Churn Analysis by Using Real-Time Machine Learning Model

Abstract: Customer churn, the loss of customers to competitors, poses a significant challenge for businesses, particularly in competitive industries such as banking and telecommunications. As a result, several customer churn analysis models have been proposed to identify at-risk customers and enable top managers to implement strategic decisions to mitigate churn and improve customer retention. Although the existing models provide top managers with promising insights for churn prediction, they rely on a batch-based training approach using fixed datasets collected at periodic intervals. While this training approach enables existing models to perform well in relatively stable environments, they, unfortunately, struggle to adapt to dynamic settings, where customer preferences shift rapidly, especially in industries with volatile market conditions, such as banking and telecom. Where, in dynamic environments, data distribution can change significantly over short periods, disabling existing models to maintain efficiency and leading to poor predictive performance, increased misclassification rates, and suboptimal decision-making by top executives, ultimately exacerbating customer churn. To address these limitations, this research proposes RCE, a real-time, continual learning-based, ensemble learning model. RCE integrates an event-driven development approach for real-time churn analysis with a replay-based continual learning mechanism to adapt to evolving customer behaviors without catastrophic forgetting, and RCE implements a stacked ensemble learning for customer churn classification. Unlike existing models, RCE continuously processes streaming data, ensuring adaptability and generalization in fast-changing environments, and providing instantaneous insights that enable decision-makers to respond swiftly to emerging risks, market fluctuations, and customer behavior changes. RCE is evaluated using the Churn Modelling benchmark dataset for European banks, achieving performance with a 95.65% accuracy; however, in dynamic environments, RCE accomplishes an average accuracy (ACC) of 86.75% and an average forgetting rate (FR) of 13.25% across tasksT_i. The results demonstrate that RCE outperforms existing models in predictive accuracy, adaptability, and robustness across multiple tasks, especially in dynamic environments. Finally, this research discusses the proposed model’s limitations and outlines directions for future improvements in real-time customer churn analysis.

Author 1: Haitham Ghallab

Author 2: Mona Nasr

Author 3: Hanan Fahmy

Keywords: Customer churn; real-time analysis; continual learning; machine learning; event-driven development; stacked ensemble learning; replay-based approach

Paper 39: Estimating Missing Data in Wireless Sensor Network Through Spatial-Temporal Correlation

Abstract: Wireless sensor networks consist of a set of smart sensors with limited memory and wireless communication capabilities. These sensors get data from the environment and send them to an application center. However, data loss has happened due to the characteristics of sensors, which negatively affect the accuracy of applications. To solve this problem, we need to estimate the missing data for applications that depend on accurate data collecting. In this study, we present an algorithm that uses the most significant historical data to estimate the missing data based on spatial and temporal correlations. In the proposed algorithm, we combine the spatial correlation by using data from the closest sensor based on the missing pattern and the temporal correlation by referring to the closest data prior to the missing instance. The experimental results demonstrate that the proposed algorithm lowers estimation errors when compared to current algorithms for a variety of missing data patterns.

Author 1: Walid Atwa

Author 2: Abdulwahab Ali Almazroi

Author 3: Eman A. Aldhahr

Author 4: Nourah Fahad Janbi

Keywords: Wireless sensor networks; missing data estimation; spatial correlation; temporal correlation

Paper 40: FPGA-Based Implementation of Enhanced DGHV Homomorphic Encryption: A Power-Efficient Approach to Secure Computing

Abstract: One new area of secure computing and privacy is homomorphic encryption (HE). An FPGA-based implementation of the HE algorithm, Enhanced DGHV, which helps real-time computation on encrypted text without disclosing the original data, is developed in this study. This research aims to focus on implementing the Enhanced DGHV Fully HE algorithm on FPGA hardware to achieve a more efficient scheme in terms of performance and security. The Xilinx Vivado tool implements the design on a Genesys 2 Kintex 7 FPGA board. While software simulation with 3.2% I/O usage, the simulation confirms a total power consumption of 3.12W (watts), highlighting successful synthesis with little resources. At 9.105W, the hardware implementation is comparable. Furthermore, an effective FPGA-based implementation confirms a method for achieving a balance between power consumption and performance while implementing the DGHV algorithm. The results show that the overall computational complexity can be reduced, and the hardware and software integration help to achieve an increased data security level for homomorphic encryption algorithms with improved efficiency.

Author 1: Gurdeep Singh

Author 2: Sonam Mittal

Author 3: Hani Moaiteq Aljahdali

Author 4: Ahmed Hamza Osman

Author 5: Ala Eldin A Awouda

Author 6: Ashraf Osman Ibrahim

Author 7: Salil Bharany

Keywords: Homomorphic encryption; cybersecurity; cryptography; DGHV; FPGA; Xilinx Vivado tool; Genesys Kintex

Paper 41: Disease Prediction from Symptom Descriptions Using Deep Learning and NLP Technique

Abstract: Accurate disease prediction from symptom descriptions is vital for improving early detection and enabling remote healthcare services, especially in the evolving landscape of digital health. Traditional diagnosis methods face significant limitations due to their reliance on structured datasets and subjective assessments, leading to delays and inefficiencies in the diagnosis process. Our strategy is to employ advanced NLP techniques such as tokenization and TF-IDF, along with DL techniques like LSTM, CNN-LSTM, and GRU, to analyze unstructured symptom data and more accurately predict diseases The study also compares two text transformation techniques (TF-IDF vectorization and tokenization) with traditional Machine Learning (ML) methods like Decision Trees to specify the best technique. Through intensive experiments on two datasets (one with 24 diseases and one with 41 diseases), the efficiency of the proposed methods is verified and the importance of using NLP and deep learning in revolutionizing healthcare is illustrated, particularly in upgrading remote diagnosis and enabling early medical intervention. The best-performing model, CNN-LSTM using tokenized text, achieved 99.90% accuracy on a 41-disease dataset, and LSTM with TF-IDF achieved 98.8% accuracy on a 24-disease dataset, outperforming or matching results from more complex models in prior studies. The findings show that combining NLP and deep learning enables accurate, efficient disease prediction, advancing remote care and early intervention in digital healthcare.

Author 1: Salmah Saad Al-qarni

Author 2: Abdulmohsen Algarni

Keywords: Natural language processing; disease prediction; machine learning; deep learning; classification

Paper 42: Digital Twin-Based Predictive Analytics for Urban Traffic Optimization and Smart Infrastructure Management

Abstract: In modern cities, urban traffic congestion remains a persistent issue that causes longer journey times, excessive fuel consumption, and environmental pollution. Traditional traffic management systems often employ static models that are insensitive to dynamic changes in urban mobility patterns in real time, which results in inefficient congestion relief. This study proposes a predictive analytics system based on digital twins to enhance smart city infrastructure management and optimize traffic flow to transcend these limitations. A Convolutional Neural Network–Gated Recurrent Unit (CNN-GRU) model is embedded at the core of the proposed system to effectively capture and learn spatial and temporal traffic patterns efficiently to enhance prediction accuracy and real-time decision-making. The scalability and robustness of the model are trained on actual urban traffic data. The system is developed and verified with Python, TensorFlow, and simulation-based digital twin platforms. The prediction capability of traffic conditions and congestion relief of the model is evidenced from the experimental results, which present a high prediction accuracy of 94.5%. Enhanced route planning, anticipatory congestion avoidance, and smart traffic signal control are some of the primary benefits. The outcome is that urban mobility has been enhanced and congestion in traffic has reduced substantially. This research contributes to the evolution of intelligent transportation systems by being the first to integrate deep learning-based predictive analytics with digital twin technology. Ultimately, the proposed framework encourages the emergence of future-oriented smart city infrastructure and the aim of sustainable city transport.

Author 1: A. B. Pawar

Author 2: Shamim Ahmad Khan

Author 3: Yousef A. Baker El-Ebiary

Author 4: Vijay Kumar Burugari

Author 5: Shokhjakhon Abdufattokhov

Author 6: Aanandha Saravanan

Author 7: Refka Ghodhbani

Keywords: Digital twin technology; traffic flow optimization; predictive analytics; smart city infrastructure; GRU-CNN hybrid model

Paper 43: Linear Correction Model for Statistical Inference Analysis

Abstract: A linear correction model based on joint independent information is proposed to optimize the statistical inference performance in high-dimensional data and small sample scenarios by integrating Fiducial inference and Bayesian posterior prediction methods. The model utilizes multi-source data features to construct a joint independent information framework, combined with an information domain dynamic correction mechanism, significantly improving parameter estimation efficiency and confidence interval coverage. Numerical simulation shows that when the sample size is 30, the posterior prediction method has a coverage rate of 0.927, approaching 95% of the theoretical value, and the coverage probability approaches the ideal level with increasing sample size. Compared with traditional methods, the model exhibits stronger adaptability and stability in high-dimensional noise covariance and dynamic data streams, providing an efficient and robust theoretical tool for statistical inference in complex data environments.

Author 1: Jing Zhao

Author 2: Zhijiang Zhang

Keywords: Linear correction model; statistical analysis; fiducial inference; numerical simulation

Paper 44: Fine-Tuning Arabic and Multilingual BERT Models for Crime Classification to Support Law Enforcement and Crime Prevention

Abstract: Safety and security are essential to social stability since their absence disrupts economic, social, and political structures and weakens basic human needs. A secure environment promotes development, social cohesion, and well-being, making national resilience and advancement crucial. Law enforcement struggles with rising crime, population density, and technology. Time and effort are required to analyze and utilize data. This study employs AI to classify Arabic text to detect criminal activity. Recent transformer methods, such as Bidirectional Encoder Representation Form Transformer (BERT) models, have shown promise in NLP applications, including text classification. Applying these models to crime prevention motivates significant insights. They are effective because of their unique architecture, especially their capacity to handle text in both left and right contexts after pre-training on massive data. The limited number of crime field studies that employ the BERT transformer and the limited availability of Arabic crime datasets are the primary concerns with the previous studies. This study creates its own X (previously Twitter) dataset. Next, the tweets will be pre-processed, data imbalance addressed, and BERT-based models fine-tuned using six Arabic BERT models and three multilingual models to classify criminal tweets and assess optimal variation. Findings demonstrate that Arabic models are more effective than multilingual models. MARBERT, the best Arabic model, surpasses the outcomes of previous studies by achieving an accuracy and F1-score of 93%. However, mBERT is the best multilingual model with an F1-score and accuracy of 89%. This emphasizes the efficacy of MARBERT in the classification of Arabic criminal text and illustrates its potential to assist in the prevention of crime and the defense of national security.

Author 1: Njood K. Al-harbi

Author 2: Manal Alghieth

Keywords: Artificial intelligence; deep learning; natural language processing; bidirectional encoder representation from transformer; crime classification; crime prevention; tweets; text classification; transformer; Arabic; X

Paper 45: Attention-Driven Hierarchical Federated Learning for Privacy-Preserving Edge AI in Heterogeneous IoT Networks

Abstract: ECG arrhythmia detection is very important in identification and management of patients with cardiac disorders. Centralized machine learning models are privacy invasive, and distributed ones poorly deal with the data heterogeneity of the devices. These challenges are responded to by presenting the edge AI an attention-driven hierarchical federated learning framework with 1-Dimensional Convolutional Neural Network (1D-CNN) - Long Short-Term Memory (LSTM) -Attention to classify arrhythmia in ECG recordings. This model includes the spatial characteristics of ECG signals and the temporal characteristics of attention maps, identifying the significant areas of the inputs and providing high interpretability and accuracy of the model. Thus, federated learning is applied to perform model training in a decentralized process through the Privacy-Preserving while the raw data remains on the edge devices. For assessment, this study utilized St. Petersburg INCART 12-lead Arrhythmia Database and Wearable Health Monitoring has given an overall classification accuracy of 96.5% with an average of AUC-ROC of 0.98 with five classes as Normal (N), Supraventricular (S), Ventricular (V), Fusion (F), and Unclassified (Q). The proposed model was created using the Python programming language with the TensorFlow framework deep learning and tested using Raspberry Pi devices to mimic edge settings. Overall, this study proves that it is possible to classify using IoT Device ECG arrhythmia reliably and securely on devices with limited resources, which will enable real-time cardiac monitoring.

Author 1: Pournima Pande

Author 2: Bukya Mohan Babu

Author 3: Poonam Bhargav

Author 4: T L Deepika Roy

Author 5: Elangovan Muniyandy

Author 6: Yousef A. Baker El-Ebiary

Author 7: V Diana Earshia

Keywords: Edge AI; federated learning; wearable health monitoring; arrhythmia; privacy-preserving; IoT device; 1DCNN-LSTM

Paper 46: Blockchain-Assisted Serverless Framework for AI-Driven Healthcare Applications

Abstract: With the advent of new sensor device designs, IoT based medical applications are increasingly being employed. This study introduces BlockFaaS: a Blockchain-assisted serverless framework that incorporates advanced AI models in latency sensitive healthcare applications with confidentiality, energy efficiency, and real-time decision-making. This framework combines the structure of AIBLOCK with dynamic sharding and zero knowledge proofs to make the framework strongly scalable with health-assured data inviolability with HealthFaaS, a serverless platform for cardiovascular risk detection. Explainable AI and federated learning models are introduced into the system to retain an equilibrium between data privacy and interpretability. All layers of communication use the Transport Layer Security protocol to ensure security. This proposed system is validated by new performance metrics such as real-time response rates and energy consumption, proving to be superior to the existing HealthFaaS and AIBLOCK technologies.

Author 1: Akash Ghosh

Author 2: Abhraneel Dalui

Author 3: Lalbihari Barik

Author 4: Jatinderkumar R. Saini

Author 5: Sunil Kumar Sharma

Author 6: Bibhuti Bhusan Dash

Author 7: Satyendr Singh

Author 8: Namita Dash

Author 9: Susmita Patra

Author 10: Sudhansu Shekhar Patra

Keywords: AIBLOCK; blockchain; healthfaas; latency optimization; serverless computing; Transport Layer Security (TLS)

Paper 47: Bridging the Gap: The Role of Education and Digital Technologies in Revolutionizing Livestock Farming for Sustainability and Resilience

Abstract: Livestock farming remains a cornerstone of global agricultural systems, contributing significantly to food security, economic development, and rural livelihoods. However, the sector is increasingly challenged by environmental degradation, inefficient practices, and socio-economic barriers. Education serves as a pivotal solution, empowering farmers with the knowledge and skills required for sustainable livestock management. This bibliometric analysis explores the intersection of livestock farming and education, analyzing research trends, thematic clusters, and collaboration patterns from 2015 to 2024 using data from the Web of Science database and VOSviewer software. The analysis identifies critical themes, such as sustainable practices, climate resilience, zoonotic disease management, and socio-economic empowerment, underscoring the transformative role of education in addressing these issues. Additionally, the integration of digital technologies, such as mobile learning platforms, precision farming tools, and blockchain-based traceability systems, enhances the accessibility and effectiveness of educational initiatives in livestock management. The findings reveal a steady growth in research on this topic, with significant academic and practical implications. Targeted educational interventions, including Farmer Field Schools and tailored training programs, are recommended to enhance productivity, promote sustainability, and foster inclusivity in the livestock sector. By integrating education with livestock farming, the study contributes to achieving Sustainable Development Goals, particularly Goals 2 (Zero Hunger) and 4 (Quality Education). This research provides a comprehensive foundation for policymakers, researchers, and practitioners to advance the integration of education in livestock farming, fostering resilience and sustainability within the sector.

Author 1: Nur Amlya Abd Majid

Author 2: Mohd Fahmi Mohamad Amran

Author 3: Muhammad Fairuz Abd Rauf

Author 4: Lim Seong Pek

Author 5: Suziyanti Marjudi

Author 6: Puteri Nor Ellyza Nohuddin

Author 7: Kemal Farouq Mauladi

Keywords: Livestock farming; sustainable agriculture; digital technologies; farmer education; climate resilience

Paper 48: Tracking Parkinson’s Disease Progression Using Deep Learning: A Hybrid Auto Encoder and Bi-LSTM Approach

Abstract: Parkinson's disease (PD) is a progressive and chronic neurodegenerative disorder characterized by motor impairment, speech deficits, and cognitive decline. Monitoring disease progression accurately and intermittently is imperative for early treatment planning and personalized intervention. In the past, conventional methods of diagnosis—clinical examination and traditional machine learning (ML) algorithms—tend to be insufficient in identifying intricate temporal behaviors of PD progress and involve frequent clinic visits. There is no cure for this disease but there are treatments. To tackle these issues, we introduce a deep learning (DL)-based approach that integrates auto encoders for feature learning with Bi-Directional Long Short-Term Memory (Bi-LSTM) networks for temporal sequence modeling. The hybrid model successfully monitors PD severity over time by learning complex patterns in the data. We measure our method with the Parkinson's Tele monitoring Dataset from the UCI Machine Learning Repository, which contains longitudinal voice recordings together with Unified Parkinson's Disease Rating Scale (UPDRS) scores—rendering it particularly well-suited for time-series analysis. Implemented in Python with Tensor Flow applies sophisticated training methods to achieve maximum performance. Experimental results affirm a dramatic improvement compared to traditional ML methods, producing an accuracy rate of 95.2%. Such high predictive power facilitates timely adjustment of treatment and improves patient management. The suggested model presents a non-invasive, scalable real-time PD monitoring solution. It aids neurologists, clinicians, and researchers by offering an AI-based platform for pre-emptive intervention. It helps patients by facilitating continuous remote monitoring, minimizing frequent clinic visits, and enhancing their quality of life.

Author 1: Sri Lavanya Sajja

Author 2: Kabilan Annadurai

Author 3: S. Kirubakaran

Author 4: TK Rama Krishna Rao

Author 5: P. Satish

Author 6: Elangovan Muniyandy

Author 7: Yahia Said

Keywords: Auto encoders; DL; Parkinson’s disease; Bi-LSTM; tele monitoring dataset

Paper 49: FB-PNet: A Semantic Segmentation Model for Automated Plant Leaf and Disease Annotation

Abstract: Semantic segmentation is an important operation in computer vision, which is generally plagued by computational resources and the time-consuming process for labor intensive of pixel-wise labeling. As a solution to this issue, the present study introduces a state-of-the-art segmentation system based on the Forward-Backward Propagated Percept Net (FB-PNet) architecture, augmented with Perception Convolution layers designed specifically for this purpose. The suggested method improves segmentation precision and processing the efficiency by capturing fine visual features and reducing some unnecessary data. The performance of the model is tested using key evaluation metrics, including Intersection over Union (IoU), Dice coefficient, Loss, Recall, and Precision. Experimental results indicate that the model works effective in segmenting leaf and disease regions in plant images without requiring full pixel-by-pixel labeling. Data augmentation techniques also greatly improve the capability of the model to handle new situations. A strong partitioning technique of the dataset allows for best performance testing, demonstrating the strength and flexibility of the model with respect to new data in the PlantVillage dataset, even without the employment of annotation masks. The innovation of this research is an efficient and scalable approach to large-scale plant leaf and disease detection, which is able to sustain precision agriculture application cases.

Author 1: P Dinesh

Author 2: Ramanathan Lakshmanan

Keywords: Semantic segmentation; forward-backward propagated percept net; intersection over union; data augmentation

Paper 50: Hybrid Sequence Augmentation and Optimized Contrastive Loss Recommendation

Abstract: To address the issues of relevance and diversity imbalance in the augmented data and the shortcomings of existing loss functions, this study proposes a recommendation algorithm based on hybrid sequence augmentation and optimized contrastive loss. First, two new data augmentation operators are designed and combined with the existing operators to form a more diversified augmentation strategy. This approach better balances the relevance and diversity of the augmented data, ensuring that the model can make more accurate recommendations when facing various scenarios. Additionally, to optimize the training process of the model, this study also introduces an improved loss function. Unlike the traditional cross-entropy loss, this loss function introduces a temporal accumulation term before calculating the cross-entropy loss, integrating the advantages of binary cross-entropy loss. This overcomes the limitation of traditional methods, which apply cross-entropy loss only at the last timestamp of the sequence, thereby improving the model's accuracy and stability. Experiments on the Beauty, Sports, Yelp, and Home datasets show significant improvements in the Hit@10 and NDCG@10 metrics, demonstrating the effectiveness of the recommendation model based on hybrid sequence augmentation and optimized contrastive loss. Specifically, the Hit metric, which reflects model accuracy, improves by 8.64%, 13.07%, 5.92%, and 19.28% respectively on these four datasets. The NDCG metric, which measures ranking quality, increases by 15.60%, 19.01%, 9.66%, and 20.31% respectively.

Author 1: Minghui Li

Author 2: Xiaodong Cai

Keywords: Recommendation algorithm; data sparsity; loss function; sequence augmentation; timestamp optimization

Paper 51: Detection of Malaria Infections Using Convolutional Neural Networks

Abstract: Malaria persists as a serious global public health threat, particularly in resource-limited regions where timely and accurate diagnosis is a challenge due to poor medical infrastructure. This study presents a comparative evaluation of three pre-trained convolutional neural network (CNN) architectures—EfficientNetB0, InceptionV3, and ResNet50—for automated detection of Plasmodium-infected blood cells using the Malaria Cell Images Dataset. The models were implemented in Python with TensorFlow and trained in Google Colab Pro with GPU A100 acceleration. Among the models evaluated, ResNet50 proved to be the most balanced, achieving 97% accuracy, a low false positive rate (1.8%) and the shortest training time (2.9 hours), making it a suitable choice for implementation in real-time clinical settings. InceptionV3 obtained the highest sensitivity (98% recall), although with a higher false positive rate (4.0%) and a higher computational demand (6.5 hours). EfficientNetB0 was the fastest model (3.2 hours), showed validation and a higher false negative rate (6.2%). Standard metrics—accuracy, loss, recall, F1-score and confusion matrix—were applied under a non-experimental cross-sectional design, along with regularization and data augmentation techniques to improve generalization and mitigate overfitting. As a main contribution, this research provides reproducible empirical evidence to guide the selection of CNN architectures for malaria diagnosis, especially in resource-limited settings. This systematic comparison between state-of-the-art models, under a single protocol and homogeneous metrics, represents a significant novelty in the literature, guiding the selection of the most appropriate architecture. In addition, a lightweight graphical user interface (GUI) was developed that allows real-time visual testing, reinforcing its application in clinical and educational settings. The findings also suggest that these models, in particular ResNet50, could be adapted for the diagnosis of other parasitic diseases with similar cell morphology, such as leishmaniasis or babesiosis.

Author 1: Luis Edison Nahui Vargas

Author 2: Mario Aquino Cruz

Keywords: Malaria diagnosis; CNN architectures; deep learning; artificial intelligence; plasmodium; clinical decision support; medical imaging

Paper 52: A Hybrid Graph Convolutional Networks (GCN)-Collaborative Filtering Recommender System

Abstract: This study proposes a hybrid recommendation system that integrates Graph Convolutional Networks (GCN) and collaborative filtering to improve the accuracy and performance of university library book recommendation systems. The goal is to develop a comprehensive evaluation method for assessing the effectiveness of recommendation algorithms in university libraries. A combination of GCN and collaborative filtering algorithms was employed to enhance recommendation accuracy. GCN was used to capture complex relationships in user data, while collaborative filtering focused on user preferences. Performance evaluation was conducted using a set of functional indicators, and the system was tested using real library data. The evaluation metrics included Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and evaluation time. The GCN-based evaluation model significantly outperformed traditional methods. It achieved a MAPE of 0.7597 and an RMSE of 0.3775, both superior to BP, CNN, and DBN algorithms. In terms of evaluation time, the GCN algorithm showed moderate performance (0.44s) compared to BP (0.32s), but better than DBN (0.87s) and CNN (0.67s). These results demonstrate the robustness and efficiency of the GCN model in predicting library recommendations. The proposed hybrid system effectively improves the accuracy and evaluation of university library recommendation systems. The GCN-based model outperformed other methods in terms of error rates and evaluation time, making it a valuable tool for enhancing personalized recommendations in library systems. Future research will focus on optimizing the computational efficiency of the GCN model.

Author 1: Qingfeng Zhang

Keywords: Graph convolutional networks; collaborative filtering; hybrid recommender systems; university library performance evaluation

Paper 53: DBSCAN Algorithm in Creation of Media and Entertainment: Drawing Inspiration from TCM Images

Abstract: This study proposes a TCM culture communication data clustering division and classification method that is based on an enhanced DBSCAN clustering algorithm and an ELM model. The objective is to address the issue of Traditional Chinese Medicine (TCM) culture communication image role product design. Firstly, for the problem of extracting feature vectors of TCM cultural communication, we analyse the path of communication role product design, design the product design scheme of TCM cultural communication image role, and extract the feature vectors of TCM cultural communication; secondly, for the problem of clustering and classifying the health data of TCM, we propose a method of clustering and classifying the health data of TCM based on the SCSO-DBSCAN clustering algorithm by combining the DBSCAN clustering algorithm with the sandcat swarm optimization algorithm. Finally, the TCM cultural dissemination data clustering classification and classification methods are tested and analyzed using TCM cultural dissemination data. This problem of TCM health data clustering classification is addressed by combining the ELM network algorithm, and a classification method of TCM cultural data dissemination based on the ELM model is proposed. The experimental results demonstrate that the method proposed in this study enhances the accuracy of TCM health data clustering division and also improves the accuracy of TCM cultural data communication classification, in comparison to other algorithms used for TCM cultural communication data clustering division and classification.

Author 1: Xiaoxiao Li

Author 2: Libo Wan

Author 3: Xin Gao

Keywords: DBSCAN algorithm; TCM cultural communication; picture character product design; sand cat swarm optimization methodology

Paper 54: GOA-WO-ML: Enhancing Internet of Things Security with Gannet Optimization and Walrus Optimizer-Based Machine Learning

Abstract: The rapid development of the Internet of Things (IoT)-based Wireless Sensor Networks (WSNs) has fueled security challenges, necessitating efficient intrusion detection approaches. The computationally intensive nature and the high-dimension data preclude the direct employment of machine learning-based Intrusion Detection Systems (IDSs). This study introduces GOA-WO-ML, a robust IDS system that integrates the Gannet Optimization Algorithm (GOA) and Walrus Optimizer (WO) for feature selection and parameter tuning in machine learning algorithms. The system is tested on the NSL-KDD dataset, indicating better cyberattack detection performance. The experimental findings suggest that GOA-WO-ML improves intrusion detection accuracy, decreases false positives, and has low computational overhead compared to traditional methods. By adopting bio-inspired methods, the proposed system successfully counteracts security issues in IoT-WSNs through efficient surveillance. Future research directions include considering deep learning improvements and real-time deployment methods in dynamic environments for further intrusion detection performance.

Author 1: Jing GUO

Author 2: Wen CHEN

Author 3: Xu ZHANG

Keywords: Internet of things; intrusion detection; machine learning; optimization

Paper 55: Efficient Task Allocation in Internet of Things Using Lévy Flight-Driven Walrus Optimization

Abstract: The rapid growth of the Internet of Things (IoT) has presented a significant challenge in efficiently managing energy-aware task distribution over heterogeneous devices. Optimizing the efficient use of resources in terms of energy consumption is critical when considering IoT device resource-constrained environments. This study proposes a new IoT task distribution resolution mechanism using an Enhanced Walrus Optimization Algorithm (EWOA). EWOA incorporates sophisticated techniques, such as Lévy flight processes and augmented exploration-exploitation, and thus is best suited to complex and dynamic IoT environments. This study proposes an EWOA to assign effective tasks considering device capability compatibility and reduced energy consumption. Simulations over benchmark IoT scenarios validate that the EWOA outperforms current approaches in terms of efficiency in terms of energy consumption, convergence, and robustness. In conclusion, improvements in minimizing energy consumption, enhancing task execution performance, and efficient use of resources in IoT networks have been emphasized significantly. In this work, the EWOA was proven to be an effective tool for IoT NP-hard optimization problem resolution and opens doors for future work in utilizing sophisticated metaheuristic algorithms for use in energy-constrained environments.

Author 1: Yaozhi CHEN

Keywords: Internet of things; energy efficiency; task scheduling; walrus; optimization

Paper 56: A Hybrid Convolutional Neural Network-Temporal Attention Mechanism Approach for Real-Time Prediction of Soil Moisture and Temperature in Precision Agriculture

Abstract: Precision Agriculture is a combination of Artificial Intelligence (AI) and the Internet of Things (IoT) to improve farming efficiency, sustainability, and overall productivity. This work presents hybrid CNN-TAM (Convolutional Neural Network–Temporal Attention Mechanism) model running on Edge AI devices for real time crop soil temperature and Soil Moisture prognosis. IoT sensors gather long term environmental data which is preprocessed to remove noise and extract meaningful spatial and temporal features. CNN can obtain spatial patterns and TAM assigns dynamic attention weights to important time steps enhancing prediction accuracy. The proposed hybrid model surpasses the conventional methods like Linear Regression, Random Forest, LSTM, and independent CNN with the lowest RMSE (1.7). Different from cloud-based deployments, the Edge AI deployment offers reduced latency, consumes lower bandwidth, and is better suited for scalability, enabling large-scale, real-time precision farming. Experimental outcome confirms enhanced real-time prediction capability allowing farmers to optimize irrigation schedules, reduce resource waste, and improve crop resilience against extreme weather conditions. This ensures sustainable resource management, conserves water and fertilizers, and enhances decision-making in agriculture. The results demonstrate the capability of AI-driven decision-support tools in present-day agriculture and presents a scalable, cost-effective and deployable solution for both small- and large-scale farms. By emphasizing data privacy, real-time processing, and low-latency inference, this research contributes to the area of precision agriculture relying on AI, addressing key challenges such as real-time analytics, unreliable connectivity, and the need for immediate on-site decision-making. The study develops an AI-powered system for intelligent farm management to support sustainable and Smart Irrigation Optimization is used for efficient agricultural practices.

Author 1: M. L. Suresh

Author 2: Swaroopa Rani B

Author 3: T K Rama Krishna Rao

Author 4: S. Gokilamani

Author 5: Yousef A.Baker El-Ebiary

Author 6: Prajakta Waghe

Author 7: Jihane Ben Slimane

Keywords: Precision agriculture; edge AI; convolutional neural network; temporal attention mechanism; smart irrigation optimization

Paper 57: Capsule Network-Based Multi-Modal Neuroimaging Approach for Early Alzheimer’s Detection

Abstract: Alzheimer’s Disease (AD) is a terminal illness affecting the human brain that leads to deterioration of cognitive function and should therefore be diagnosed as early as possible. The goal of this work is to come up with a precise and interpretable diagnostic model for the early diagnosis of Alzheimer's Disease (AD) based on multi-modal neuroimaging data. Current deep learning models such as Convolutional Neural Networks (CNNs) are limited in that they lose spatial hierarchies in 3D medical images, which inhibits classification performance and interpretability. To overcome this, in this work, we introduce a new 3D Capsule Network (3D-CapsNet) framework that captures spatial relations more effectively with dynamic routing and pose encoding to improve volumetric neuroimaging data analysis. Our approach has three principal phases: extensive pre-processing of MRI and PET scans such as skull stripping, intensity normalization, and motion correction; feature extraction through the 3D-CapsNet model; and multi-modal classification based on fusion. We used the Alzheimer's Classification dataset from Kaggle for training and testing. The model is implemented in the Python platform with TensorFlow and Keras libraries incorporating 3D CNN operations along with capsule layers to extract fine-grained features of AD-affected brain areas such as the hippocampus and entorhinal cortex. Experimental results show that our model reaches a very high classification accuracy of 92%, which is higher than the conventional architectures VGG-16, ResNet-50, and DenseNet-121 in accuracy, precision, recall, F1-score, and AUC-ROC. This strategy is helpful to clinicians and medical researchers because it gives them a non-invasive, interpretable, and trustworthy tool for diagnosing and monitoring various stages of AD (Non-Demented, Very Mild, Mild, and Moderate). It sets the stage for real-time clinical integration and future studies in monitoring disease progression over time.

Author 1: Kabilan Annadurai

Author 2: A Suresh Kumar

Author 3: Yousef A.Baker El-Ebiary

Author 4: Sachin Upadhye

Author 5: Janjhyam Venkata Naga Ramesh

Author 6: K. Lalitha Vanisree

Author 7: Elangovan Muniyandy

Keywords: Alzheimer’s detection; 3d-capsule networks; multi-modal neuroimaging; deep learning in healthcare; early diagnosis and classification

Paper 58: Neuro-Symbolic Reinforcement Learning for Context-Aware Decision Making in Safe Autonomous Vehicles

Abstract: Autonomous vehicles need to be equipped with smart, understandable, and context-aware decision-making frameworks to drive safely within crowded environments. Current deep learning approaches tend to generalize poorly, lack transparency, and perform inadequately in dealing with uncertainty within dynamic city environments. Towards overcoming these deficiencies, this study suggests a new hybrid approach that combines Neuro-Symbolic reasoning with a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture, together with a Deep Q-Network (DQN) for learning through reinforcement. The model employs symbolic logic to enforce traffic regulations and infer context while relying on CNN for extracting spatial features and LSTM for extracting temporal dependencies in vehicle motion. The system is trained and tested using the Lyft Level 5 Motion Prediction dataset, which emulates varied and realistic driving scenarios in urban environments. Enforced on the Python platform, the new framework allows autonomous cars to generate rule-adherent, strong, and explainable choices under diverse driving scenarios. Neuro-symbolic combination is more robust for learning as well as explainability, whereas reinforcement improves long-term rewards regarding safety and efficiency. The experiment shows that the model provides high accuracy of 98% on scenario-based decision-making problems in contrast to classical deep learning models used in safety-critical routing. This work is advantageous to autonomous vehicle manufacturers, smart mobility system developers, and urban planners by providing a scalable, explainable, and reliable AI-based solution for future transportation systems.

Author 1: Huma Khan

Author 2: Tarunika D Chaudhari

Author 3: Janjhyam Venkata Naga Ramesh

Author 4: A. Smitha Kranthi

Author 5: Elangovan Muniyandy

Author 6: Yousef A.Baker El-Ebiary

Author 7: David Neels Ponkumar Devadhas

Keywords: Autonomous vehicles; neuro-symbolic learning; Deep Q-Network (DQN); CNN-LSTM architecture; context aware

Paper 59: Quantum-Assisted Variational Deep Learning for Efficient Anomaly Detection in Secure Cyber-Physical System Infrastructures

Abstract: The aim of the current study is to propose a Quantum-Assisted Variational Autoencoder (QAVAE) model capable of efficiently identifying anomalies in high-dimensional, time-series data produced by cyber-physical systems. The existing approaches to machine learning have some limitations when recording temporal interactions and take substantial time to run with many attributes. To meet all of these challenges, this study aims at proposing a quantum-assisted approach to anomaly detection using the potential of a Quantum-Assisted Variational Autoencoder (QAVAE). The general goal of this research is to optimize anomaly detection systems using consummate deep learning quantum computing models. According to the QAVAE framework, variational inference is employed for learning latent representations of time series data; besides, quantum circuits are utilized for enhancing the capacity of the model and its generalization capability. This work was accomplished using Python programming language, and the analysis was carried out using TensorFlow Quantum. The QAVAE model demonstrates the highest accuracy of 95.2%, indicating its strong capability in correctly identifying both anomalous and normal instances. So, it can learn well from the data and keep stable in the evaluation process, which will make it suitable for real-time anomaly detection in dynamic environments. In conclusion, the QAVAE model brings a reasonable approach and solution for anomaly detection that is accurate in identifying and scalable too. Utilizing the HAI, the dataset achieved a high detection accuracy of 95.2%. Further research has to be dedicated to its application to quantum computing architecture as well as to modifications that allow for its use on multi-variable actual-life data.

Author 1: Nilesh Bhosale

Author 2: Bukya Mohan Babu

Author 3: M. Karthick Raja

Author 4: Yousef A.Baker El-Ebiary

Author 5: Manasa Adusumilli

Author 6: Elangovan Muniyandy

Author 7: David Neels Ponkumar Devadhas

Keywords: Quantum variational circuits; cyber-physical system security; hybrid quantum-classical algorithms; anomaly detection framework; quantum machine learning optimization

Paper 60: EJAIoV: Enhanced Jaya Algorithm-Based Clustering for Internet of Vehicles Using Q-Learning and Adaptive Search Strategies

Abstract: The Internet of Vehicles (IoV) is an indispensable part of contemporary Intelligent Transportation Systems (ITS), providing efficient vehicle-to-everything (V2X) communication. Nevertheless, high mobility and consequent topological changes in IoV networks create overwhelming difficulties in establishing and maintaining stable and effective communication. In this work, we introduce the Enhanced Jaya Algorithm for IoV (EJAIoV), an optimized clustering algorithm using optimization to develop stable and long-term clusters in IoV scenarios. EJAIoV uses efficient random initialization with three scrambling strategies to produce diverse, high-quality solutions. Q-learning selection between three neighborhood operators enhances local search effectiveness by incorporating a segmented operator. In addition, an adaptive search balance strategy adjusts solution updating dynamically to avoid premature convergence and optimize the exploration procedure. Simulation experiments show that EJAIoV outperforms existing clustering algorithms, achieving up to 31.5% improvement in cluster lifetime and 28.2% reduction in the number of clusters across various node densities and grid sizes.

Author 1: Jinchuan LU

Keywords: Internet of vehicles; clustering; Jaya algorithm; Q-learning; optimization

Paper 61: CT Imaging-Based Deep Learning System for Non-Small Cell Lung Cancer Detection and Classification

Abstract: About 85% of all occurrences of lung cancer are classified as Non-Small Cell Lung Cancer (NSCLC), making it a serious worldwide health concern. For better treatment results and patient survival, NSCLC must be detected early and accurately. This research presents an advanced Deep Learning-enabled Lung Cancer Detection and Classification System (LCDCS) aimed at significantly improving diagnostic precision and operational efficiency. Emerging technologies such as artificial intelligence and multi-level convolutional neural networks (ML-CNN) are increasingly being leveraged in CT imaging-based deep learning systems for accurate detection. The outlined framework leverages a multi-layer convolutional neural network to effectively analyse CT scan images and accurately classify lung nodules. Tomek link and Adaptive Synthetic Sampling (ADASYN) are used in a novel way to balance data, address class imbalance, and guarantee strong model performance. Deep learning with a CNN model is utilized to derive features, and the SoftMax function is applied for multi-class classification. Thorough evaluation on datasets like the LUNA16 dataset demonstrates that the system surpasses earlier models and data balancing techniques in accuracy, yielding a training accuracy of 95.8% and a validation accuracy of 96.9%. The findings demonstrate the potential of the suggested method as a trustworthy diagnostic instrument for the prompt identification of lung cancer. The study emphasizes on how crucial it is to combine deep learning architectures with sophisticated data balancing techniques to overcome medical imaging difficulties and raise diagnostic accuracy. Future research attempts to investigate real-time deployment in clinical settings and expand the system's capability to encompass more cancer types.

Author 1: Devyani Rawat

Author 2: Sachin Sharma

Author 3: Shuchi Bhadula

Keywords: Artificial intelligence; NSCLC; ML-CNN; ADASYN; tomek link

Paper 62: Intelligent Identification of Pile Defects Based on Improved LSTM Model and Wavelet Packet Local Peaking Method

Abstract: With the continuous expansion of building scale, the structural safety of foundation piles, as key load-bearing components, has received increasing attention. To improve the defect recognition ability under complex working conditions, this study first uses the whale optimization algorithm to perform hyperparameter optimization on the long short-term memory network model, achieving efficient classification of the defect and non-defect samples. Subsequently, the signals identified as having defects are subjected to wavelet packet decomposition to extract multi-scale energy features, and combined with the local peak finding method to accurately locate key reflection peaks, achieving further identification of defect types. The results showed that the classification accuracy, recognition precision, recall rate, and F1 value of the new method were the highest at 96.7%, 95.16%, 93.87%, and 94.51%, respectively, and the average recognition time was the shortest at 0.97 seconds. Especially for the defect identification errors of drilled cast-in-place piles and prefabricated piles, the lowest were 0.19 and 0.23, and the lowest complexity could reach 65.28%, demonstrating high precision and stability in defect identification. This model has strong robustness and accuracy in various types of defect scenarios, and has good generalization ability and engineering application potential, which can provide certain technical references for the construction monitoring of road and bridge engineering in the future.

Author 1: Xiaolin Li

Author 2: Xinyi Chen

Keywords: Foundation pile; defect identification; LSTM; WOA; WPT; LPS

Paper 63: Internet of Things-Driven Safety and Efficiency in High-Risk Environments: Challenges, Applications, and Future Directions

Abstract: The Internet of Things (IoT) is a technology that can bring about significant change in several areas, especially in high-risk situations such as industrial environments and health and safety contexts. This research study has examined many IoT applications within domains and identified their importance in improving risk management and operational efficiency strategies. IoT enables sensor networks, wearable devices, and remote monitoring systems with edge computing capabilities. Thus, it allows real-time monitoring, early threat detection, and predictive maintenance. Data analytics technologies make it easier to capture valuable information that stakeholders can use to make informed decisions and optimize workflows to improve performance. Despite the transformational promises of IoT, there are still some problems. These include security vulnerabilities, interoperability concerns, and extensive training programs. Addressing these challenges offers the opportunity to create innovative, resourceful collaboration in developing robust IoT solutions to accommodate the requirements of hazardous environments. In the coming times, further growth of IoT and integration with the latest technologies like 5G and robotics promise new ways to ensure safety and efficiency in operations. Within this study, we emphasize the role of IoT as an enabling factor in transforming dangerous areas into safe and efficient zones, assuring our readers on the safety benefits of IoT. It also provides a general perspective towards potential future research and development directions.

Author 1: Hua SUN

Keywords: Internet of things; high-risk environments; safety; operational efficiency; data analytics

Paper 64: ECOA: An Enhanced Chimp Optimization Algorithm for Cloud Task Scheduling

Abstract: Effective scheduling of tasks is a key concern in cloud computing because it considerably affects system functionality, resource usage, and execution efficiency. The present study proposes an Enhanced Chimp Optimization Algorithm (ECOA) to address such problems by overcoming the disadvantages of traditional scheduling methods. The proposed ECOA combines three innovative components: 1) the highly disruptive polynomial mutation enhances population diversity, 2) the Spearman rank correlation coefficient promotes the refinement of inferior solutions, and 3) the beetle antennae operator facilitates more efficient local exploitation. These changes significantly enhance the equilibrium between exploration and exploitation, decrease the chance of premature convergence, and are a better solution. Extensive experiments on benchmark datasets prove that ECOA outperforms traditional algorithms concerning makespan, imbalance degree, and resource utilization. The obtained results confirm that the proposed ECOA has excellent potential for better performance in task scheduling in dynamic and large-scale cloud environments, as it represents a promising optimization solution for complex problems in cloud computing.

Author 1: Yue WANG

Keywords: Cloud computing; task scheduling; resource utilization; chimp optimization

Paper 65: PSOMCD: Particle Swarm Optimization Algorithm Enhanced with Modified Crowding Distance for Load Balancing in Cloud Computing

Abstract: Effective load balancing in cloud computing architectures is crucial towards enhancing resource utilization, response times, and stability in the system. The present study proposes a new strategy with a Particle Swarm Optimization algorithm enhanced with Modified Crowding Distance (PSOMCD) to tackle task scheduling among Virtual Machines (VMs) in dynamic scenarios. The traditional PSO algorithm is supplemented by an enhanced crowding distance mechanism by PSOMCD to improve diversity in decision spaces and convergence to optimal solutions. The multi-objective fitness function addresses principal challenges in cloud computing, including load distribution, energy consumption, and throughput optimization. The performance of the algorithm is demonstrated in simulations, comparing its performance with other optimization techniques available in the literature. Results prove that PSOMCD provides better task allocation, improved load balancing, and decreased energy usage, thus effectively managing resources in dynamic and heterogeneous cloud ecosystems.

Author 1: Bolin ZHOU

Author 2: Jiao GE

Author 3: RuiRui ZHANG

Keywords: Cloud computing; load balancing; particle swarm optimization; crowding distance; task allocation

Paper 66: Hybrid Meta-Heuristic Algorithm for Optimal Virtual Machine Migration in Cloud Computing

Abstract: Virtual Machine (VM) migration is one of the most important features of cloud computing for resource utilization optimization, energy minimization, and quality of service enhancement. Existing migration solutions, however, suffer from excessive migration overhead, energy inefficiency, and ineffective allocation of resources. This study proposes a novel hybrid meta-heuristic algorithm through the integration of Particle Swarm Optimization (PSO) and Seahorse Optimization (SHO) to address the drawbacks. The proposed PSOSHO algorithm takes advantage of the global exploration capability of PSO and the adaptive exploitation feature of SHO and provides a sound solution for VM migration in dynamic cloud computing environments. Extensive simulation experiments were conducted for a different number of cloud tasks, and the results demonstrated that PSOSHO significantly outperforms existing algorithms. Specifically, it achieves improvements of up to 54% in load factor, 60% in migration count, 48% in migration cost, 7% in energy consumption, 27% in resource availability, and 37% in computation time. These results confirm the effectiveness and robustness of the proposed methodology for optimal VM migration and resource management in virtualized cloud computing infrastructures.

Author 1: Hongkai LIN

Keywords: Cloud computing; virtualization; migration; particle swarm optimization; seahorse optimization

Paper 67: Real-Time Emotion Recognition in Psychological Intervention Methods

Abstract: In the context of mental health, this study aims to develop a real-time emotion-focused facial recognition system based on psychological intervention methods. It uses a convolutional neural network (CNN) base and is trained with the FER2013 dataset, which consists of 35,887 facial images classified into seven basic emotions. Through normalisation, data augmentation, and training in TensorFlow and Keras, the model achieved 92.3% accuracy in a pilot test with 1,000 images, achieving an F1 score of 0.92, precision of 0.93, and recall of 0.91. Subsequently, when scaled to 71,774 images, it maintained robust performance with an overall accuracy of 77.5%. Emotions such as happiness (0.83), surprise (0.80), and neutrality (0.85) were recognised with greater accuracy, while K-means analysis was applied to cluster emotional patterns in a visually interpretable way. Complementing the technical architecture, a user-friendly graphical interface was designed for psychology professionals, allowing clear visualisation of the detected emotions with a latency of just 150 milliseconds per image. Overall, this proposal represents a significant advance toward more interactive, personalised, and efficient therapies, without requiring a complex technological infrastructure. Future studies recommend exploring different multimodal signals and increasing the use of convolutional layers to improve the quality of results and data efficiency.

Author 1: Sebastián Ramos-Cosi

Author 2: Daniel Yupanqui-Lorenzo

Author 3: Meyluz Paico-Campos

Author 4: Claudia Marrujo-Ingunza

Author 5: Ana Huamaní-Huaracca

Author 6: Maycol Acuña-Diaz

Author 7: Enrique Huamani-Uriarte

Keywords: Facial recognition; real-time; methods; psychological interventions

Paper 68: Artificial Intelligence-Driven Physical Simulation and Animation Generation in Computer Graphics

Abstract: This study explores an expression synthesis algorithm anchored in Generative Adversarial Networks (GAN) with attention mechanisms, achieving enhanced authenticity in facial expression generation. Evaluated on the MUG and Oulu-CASIA datasets, our method synthesizes six expressions with superior clarity (96.63±0.26 confidence for neutral expressions) and smoothness (SSIM >0.92 for video frames), outperforming StarGAN and ExprGAN in detail preservation and temporal stability. The proposed model demonstrates significant advantages in realism and identity retention, validated through quantitative metrics and comparative experiments.

Author 1: Fei Wang

Keywords: GAN; computer graphics; expression synthesis; animation generation

Paper 69: Power Line Fault Detection Combining Deep Learning and Digital Twin Model

Abstract: To address the issue of inadequate diagnosis of power line faults, an automated power line fault diagnosis technology is put forward. In this context, the research leverages the object detection algorithm YOLOv5 to construct a fault diagnosis model and enhances its anchor box loss function. In addition, the study introduces digital twin models for fault point localization, and improves the recognition model by introducing GhostNet and attention mechanism, thereby enhancing the diagnostic performance of the technology in multi-objective scenarios. In the performance test of the loss function, the improved loss function performs the best in both regression loss and intersection over union ratio, with the average loss value and intersection over union ratio being 125 and 0.986, respectively. In multi-scenario fault diagnosis, the research model performs the best in accuracy and model loss, with values of 0.986 and 0.00125, respectively. In multi-scenario fault diagnosis, such as abnormal heating detection, when the number of targets is 4, the relative error of the research model is 0.86%, which is better than similar models. Finally, in the testing of frame rate recognition and diagnostic time, the research model shows excellent performance, surpassing similar technologies. The technology proposed by the research has good application effects. This study provides technical support for the construction of power informatization and line maintenance.

Author 1: Siyu Wu

Author 2: Xin Yan

Keywords: YOLOv5; route; fault diagnosis; digital twin; loss function

Paper 70: Maximizing Shift Preference for Nurse Rostering Schedule Using Integer Linear Programming and Genetic Algorithm

Abstract: This study explores how scheduling methods can support work-life balance and overall job satisfaction by considering the preferences of the nursing staff. Creating a nurse rostering schedule that maximizes staff preferences for working shifts, off days, and hospital demands was the main goal. A Google Form that was distributed to the nursing staff is used to gather preference data. With the help of the LPSolve IDE, an integer linear programming (ILP) technique is used for the first datasets, and the Flexible Shift Scheduling System is utilized to facilitate the use of a genetic algorithm approach for the second dataset. The first dataset's result reveals that the proposed schedule's preference weight is 205.8 (73.35%), indicating an increase of 46.24 (16.48%) over the current schedule's 159.56 (56.87%) preference weight. According to the results of the second dataset, the preference weight for the current schedule is 589 (62.98%), whereas the preference weight for the proposed schedule is 619.2 (66.21%), indicating a 30.2 (3.23%) increase. This demonstrates that both proposed schedules have higher preference weight values than the current schedule, which satisfies the study's primary goal of optimizing staff preferences. The genetic algorithm is used in the second dataset since it has a high complexity problem and can produce a near-optimal solution. Flexible Shift Scheduling System generates quicker and easier schedules compared to manual schedules. This study emphasizes the importance of including nurse staff preferences into consideration when creating nurse rostering schedule procedures to support a happier and more engaged nursing team.

Author 1: Siti Noor Asyikin Binti Mohd Razali

Author 2: Thesigan Achari A/L Tamilarasan

Author 3: Batrisyia Binti Basri

Author 4: Norazman bin Arbin

Keywords: Nurse rostering schedule; schedule optimization; metaheuristic techniques; complex scheduling; integer linear programming; genetic algorithm; shift; and off-day preference maximization

Paper 71: The Innovative Design System of Traditional Embroidery Patterns Based on Computer Linear Classifier Intelligent Algorithm Model

Abstract: This research introduces an innovative system for designing traditional embroidery patterns utilizing a computer-based linear classifier intelligent algorithm. The system achieves efficient classification and recognition of embroidery pattern features by employing the Fisher linear discriminant analysis technique, thus enabling the intelligent and innovative creation of designs. Additionally, the system encompasses the design of classification algorithms for embroidery patterns and incorporates interactive tools along with embroidery systems, offering designers a user-friendly platform for pattern creation. In the system design, Fisher linear discriminant analysis algorithm is used to classify the feature vectors of embroidery patterns to ensure that the features of each type of pattern are accurately extracted and identified. The model simulation verifies the algorithm's effectiveness through multiple iterations, and the results show that the system has significantly improved the classification accuracy of embroidery patterns and the efficiency of innovative design. Accurate data analysis shows that the classification accuracy of the system in different types of embroidery patterns reaches more than 95%, and user satisfaction is improved by 20%.

Author 1: Xiao Bai

Keywords: Fisher linear discriminant analysis; embroidery pattern; interactive tool; embroidery interactive system

Paper 72: Support Vector Machine with Rule Extraction to Improve Diabetes Prediction Using Fuzzy AHP-Sugeno and Nearest Neighbor

Abstract: Diabetes is one of the most prevalent chronic diseases globally, with significant mortality and morbidity rates. Early and accurate diagnosis plays a critical role in managing and mitigating its impact. However, achieving high diagnostic accuracy while ensuring interpretability remains a key challenge in medical machine learning applications. This paper proposes an interpretable and accurate hybrid framework for diabetes prediction that integrates Support Vector Machine Rule Extraction (SVMRE), Fuzzy Analytic Hierarchy Process (Fuzzy AHP), and Sugeno fuzzy inference. The primary objective of this study is to enhance prediction accuracy while enabling the extraction of meaningful and explainable decision rules derived from SVM models. To address the black-box nature of traditional SVM models, fuzzy rules are extracted and embedded into a Sugeno fuzzy inference system. Attribute importance is quantified through Fuzzy AHP based on expert consultation, ensuring medically relevant decision-making. Furthermore, to overcome rule redundancy and complexity, the coefficient of variation is computed for each rule and optimized using a Nearest Neighbor (NN) approach, which clusters rules with adjacent variation values. The proposed framework is evaluated using a real-world diabetes dataset from Sylhet, Bangladesh. It achieves a prediction accuracy of 84.62 per cent, outperforming several conventional methods. Compared to other competitive approaches found in recent literature, such as fuzzy grey wolf optimization and neuro-fuzzy systems, our method demonstrates superior balance between interpretability, computational efficiency, and classification performance. This study confirms that integrating rule-based learning, fuzzy expert systems, and statistical optimization provides a robust and interpretable approach for diabetes prediction. The framework aligns with Sustainable Development Goal 3 (SDG 3) by promoting early detection and decision support for non-communicable diseases in healthcare systems.

Author 1: Muhammadun

Author 2: Baity Jannaty

Author 3: Rajermani Thinakaran

Author 4: Taufik Rachman

Keywords: SVM; Fuzzy AHP; rule extraction; diabetes; coefficient of variation; fuzzy Sugeno; SDG 3

Paper 73: Spatiotemporal Modeling of Foot-Strike Events Using A0-Mode Lamb Waves and 2D Wave Equations for Biomechanical Gait Analysis

Abstract: This study introduces a physics-based framework for modeling human running biomechanics by interpreting foot-strike events as point-source excitations generating radially propagating wavefronts, akin to A0-mode Lamb waves, in a cylindrical coordinate system. Using a two-dimensional damped wave equation solved via finite-difference methods, we simulate spatiotemporal displacement fields and compare the outcomes with real-world gait kinematic and kinetic data. Our approach performs a parameter sweep of excitation frequency and amplitude to identify configurations closely replicating biomechanical signals associated with different running profiles and injury states. Unlike traditional machine learning approaches, our model leverages physical wave dynamics for simulation-validation matching, enabling interpretable identification of anomalies and potential injury risks. The results reveal distinctive wave propagation patterns between injured and non-injured runners, supporting the viability of wave-based modeling as a diagnostic and analytic tool in sports biomechanics. This work opens a novel direction for physics-informed, data-driven hybrid methods in gait analysis and injury prevention.

Author 1: Tajim Md. Niamat Ullah Akhund

Author 2: Waleed M. Al-Nuwaiser

Author 3: Md. Sumon Reza

Author 4: Watry Biswas Jyoty

Keywords: Biomechanics; foot-strike modeling; lamb waves; wave equation; gait analysis; Internet of Things (IoT); Human-Computer Interaction (HCI)

Paper 74: Integrating AI in Ophthalmology: A Deep Learning Approach for Automated Ocular Toxoplasmosis Diagnosis

Abstract: Background: Ocular Toxoplasmosis, a leading cause of Posterior Uveitis, demands timely diagnosis to prevent vision loss. Manual retinal image analysis is labor-intensive and variable, while existing Deep Learning models often fail to balance local details and global context in Medical Image Classification. Objective: I propose RetinaCoAt, a Hybrid Deep Learning Model based on the CoAtNet Architecture, for Automated Diagnosis of Ocular Toxoplasmosis, integrating local and global features in Retinal Image Analysis. Methods: RetinaCoAt combines Convolutional Neural Networks (CNNs) for local pathological pattern detection with Transformer Models using multi-head self-attention for global context. Enhanced by residual connections and optimized tokenization, it was trained on 3,659 retinal images (healthy vs. unhealthy) and benchmarked against VGG16, CNNs, and ResNet. Results: RetinaCoAt achieved 98% accuracy in Medical Image Classification, outperforming VGG16 (96.87%), CNNs (95%), and ResNet (93.75%), due to its robust CNN-Transformer synergy. Conclusion: RetinaCoAt advances Automated Diagnosis of Ocular Toxoplasmosis and Posterior Uveitis, with potential for broader retinal pathology detection.

Author 1: Bader S. Alawfi

Keywords: Ocular Toxoplasmosis; Posterior Uveitis; deep learning; automated diagnosis; CNNs; transformer models; CoAtNet architecture; retinal image analysis; medical image classification; hybrid deep learning models

Paper 75: Behavioural Analysis of Malware by Selecting Influential API Through TF-IDF API Embeddings

Abstract: The constant threat of malware makes studying its behavior an ongoing task. Malware identification and clas-sification challenges can be solved better by analyzing software behaviorally rather than using conventional hashcode-based signatures. API sequence represents the behavior of any program when collected during its execution. Considering API sequences gathered while the malware was being executed in controlled conditions, this report addresses the issue of choosing influential APIs for malware. The suggested feature selection method Select API in this research selects key features, i.e., significant APIs, that can better classify malware using TF-IDF API embeddings. Two machine learning models, Random Forest, which ensemble several estimators implicitly, and Support Vector Classifier, a standard non-linear model, are trained and evaluated to validate the importance of the chosen APIs. The proposed API selection methodology, called SelectAPI, has shown promising results. It achieves accuracy, macro-avg precision-score, macro-avg recall-score, and macro-avg F1-score of 0.76, 0.77, 0.76, and 0.76, respectively. This method focuses on selecting influential APIs and has resulted in significantly improved performance on the open-benchmark multiclass dynamic-API-Sequence based malware dataset, MAL-API-2019. These results surpass the previously best-known accuracy value of 0.60 and reported F1-Score of 0.61.

Author 1: Binayak Panda

Author 2: Sudhanshu Shekhar Bisoyi

Author 3: Sidhanta Panigrahy

Keywords: Malware analysis; behavioural analysis; API sequence; multiclass malware; TF-IDF; API embeddings

Paper 76: A Layered Security Perspective on Internet of Medical Things: Challenges, Risks, and Technological Solutions

Abstract: The Internet of Medical Things (IoMT) refers to smart devices that are used in their transformation of the healthcare sector with continuous monitoring in real time, remote diagnostics as well as real time data exchange. nevertheless such systems are being targeted by a number of challenges like data breaches, unauthorized users and service interruptions. The study uses the PRISMA 2020 method and analyzes 25 peer-reviewed articles that were published between 2020 and 2025. Security risks are identified and mapped on the IoMT architecture’s perception, network, application and cloud layers. One of the key findings was confirming the fact that blockchain based identity management, algorithmic lightweight cryptographic protocol, and Artificial Intelligence(AI) driven intrusion detection systems can potentially address these risks. However, these areas are still limited in terms of interoperability, resource efficiency, and there are no solutions against the emerging quantum threats. A number of countermeasures achieved almost perfect detection accuracy over 98%, leading to increased security for IoMT systems. In order to solve the above issues, the framework, TrustMed-IoMT, is introduced to integrate blockchain-based identity management, intelligent intrusion detection and encryption that is safe against quantum attacks.

Author 1: Ziad Almulla

Author 2: Hussain Almajed

Author 3: M M Hafizur Rahman

Keywords: IoMT; security risks; challenges; healthcare IoT; countermeasures; TrustMed-IoMT

Paper 77: Predictive Maintenance Based on Deep Learning: Early Identification of Failures in Heavy Machinery Components

Abstract: Deep learning-based predictive maintenance is a key strategy in industry to prevent unexpected failures, reduce downtime, and improve operational safety. This study presents an advanced approach for early fault detection in heavy machinery components using image analysis, focusing on four critical defect types: hose wear, piston failure, corrosion, and moisture. To this end, three state-of-the-art object detection models were implemented and compared: YOLOv11, RT-DETR, and YOLO-World. The dataset consists of images captured in real-life industrial environments exhibiting variations in lighting, texture, and material degradation. A manual preprocessing and annotation process was applied to improve training quality. Model performance was evaluated using key metrics such as the precision-recall (PR) curve and the confusion matrix to determine the most efficient technique for real-time fault detection. Experimental results show that YOLOv11 achieves the highest overall accuracy, with an mAP@0.5 of 83.8%, followed by YOLO-World at 82.4% and RT-DETR at 80.3%. In terms of efficiency, YOLO-World offers a balance between accuracy and detection speed, while RT-DETR shows stable performance but lower accuracy for certain defect types. These findings confirm that deep learning-based detection models enable the rapid and accurate identification of industrial defects, facilitating the implementation of predictive maintenance strategies.

Author 1: Pablo Cabrera Melgar

Author 2: Luis Hilasaca Chambi

Author 3: Raul Sulla Torres

Keywords: Predictive maintenance; deep learning; fault detection; artificial intelligence

Paper 78: Enhancing Topic Interpretability with ChatGPT: A Dual Evaluation of Keyword and Context-Based Labeling

Abstract: Accurate topic labeling is essential for structuring and interpreting large-scale textual data across various domains. Traditional topic modeling methods, such as Latent Dirichlet Allocation (LDA), effectively extract topic-related keywords but lack the capability to generate semantically meaningful and contextually appropriate labels. This study investigates the integration of a large language model (LLM), specifically ChatGPT, as an automatic topic label generator. A dual evaluation frame-work was employed, combining keyword-based and context-based assessments. In the keyword-based evaluation, domain experts reviewed ChatGPT-generated labels for semantic relevance using LDA-derived keywords. In the context-based evaluation, experts rated the alignment between ChatGPT-assigned topic labels and actual content from representative sample posts. The findings demonstrate strong agreement between AI-generated labels and human judgments in both dimensions, with high inter-rater reliability and consistent contextual relevance for several topics. These results underscore the potential of LLMs to enhance both the coherence and interpretability of topic modeling outputs. The study highlights the value of incorporating context in evaluating automated topic labeling and affirms ChatGPT’s viability as a scalable, efficient alternative to manual topic interpretation in research, business intelligence, and content management systems.

Author 1: Mashael M. Alsulami

Author 2: Maha A. Thafar

Keywords: Automatic label generation; topic modeling; Large Language Models (LLMs); topic labeling; semantic relevance

Paper 79: Detecting Hate Speech Targeting Protected Groups in Arabic Using Hypothesis Engineering and Zero-Shot Learning with Ground Validation via ChatGPT

Abstract: Automatic detection of hate speech in low-resource languages presents a persistent challenge in natural language processing, particularly with the rise of toxic discourse on social media platforms. Arabic, characterized by its rich morphology, dialectal variation, and limited annotated datasets, is underrep-resented in hate speech research, especially regarding content targeting marginalized and protected groups. This study proposes a zero-shot learning approach that leverages Natural Language Inference (NLI) models guided by carefully engineered hypotheses in native Arabic to detect hate speech against protected groups, such as women, immigrants, Jews, Black people, transgender individuals, gay people, and people with disabilities. We formulated nine different Arabic hypothesis groups and employed a zero-shot XNLI model with a baseline embedding-based model, incorporating preprocessing techniques on the HateEval Arabic dataset. The results indicate that the XNLI model achieves up to 80% accuracy in detecting targeted hate speech, significantly out-performing baseline models. Furthermore, a real-world validation using GPT-3 via the ChatGPT interface achieved 54% accuracy in zero-shot conversational settings. These findings highlight the importance of hypothesis design and linguistic preprocessing in zero-shot hate speech detection, particularly in low-resource and culturally nuanced languages offering a scalable and culturally aware solution for moderating harmful content in Arabic online spaces.

Author 1: Ahmed FathAlalim

Author 2: Yongjian Liu

Author 3: Qing Xie

Author 4: Alhag Alsayed

Author 5: Musa Eldow

Keywords: Hate speech detection; low resource Arabic language; zero-shot learning; natural language processing; ChatGPT; transfer learning; online safety

Paper 80: Semantic and Fuzzy Integration: A New Approach to Efficient and Flexible Querying of Relational Databases

Abstract: Data are “gold mines” that must be processed and interpreted quickly and efficiently to be useful. Thus, flexible queries continue to attract considerable attention. Several works have been proposed that allow users to perform flexible queries on relational databases. Most are related to fuzzy logic, which showed its performance in handling fuzziness in scalar values, but non-scalar values are still a more complex task. To solve this drawback of fuzzy logic, we propose using ontologies to establish the semantic relationships between the domain elements of a queried attribute. Moreover, we present the architecture of a new system that combines both techniques to allow users to write and execute queries in a flexible way where the criteria are not only exact but can also be fuzzy or semantic, and they may also include accomplishment degrees. Furthermore, the proposed system uses a new fast methodology for handling fuzzy queries, which has shown its great efficiency in accelerating the execution of fuzzy queries. Data mining techniques are used to assist users in defining their fuzzy understanding. The developed system has a user-friendly interface to assist users in managing their fuzzy preferences and semantic preferences. Finally, we have proven the performance of our system by conducting a set of experiments in different areas. We have also provided a qualitative and quantitative comparison with flexible query systems, which are documented in the literature.

Author 1: Rachid Mama

Author 2: Mustapha Machkour

Keywords: Relational databases; fuzzy logic; ontologies; flexible queries; user interface

Paper 81: MRI Brain Tumor Image Enhancement Using LMMSE and Segmentation via Fast C-Means

Abstract: Brain MRI imaging revolutionizes tumor diagnosis, yet noise frequently obscures the images, complicating precise tumor identification and segmentation. This paper presents a comprehensive pipeline for brain MRI enhancement and tumor segmentation. The proposed method integrates Wavelet Packet Transform (WPT) and Linear Minimum Mean Square Error (LMMSE) filtering for effective noise reduction, combined with morphological operations for contrast enhancement. For segmentation, Fast C-Means clustering is employed, with the number of clusters automatically determined from histogram peaks. The tumor cluster is selected based on the highest centroid intensity and further refined by morphological operations to accurately delineate tumor borders. The approach is evaluated on the BraTS 2021 dataset, subject to Rician, Gaussian, and salt-and-pepper noise with intensities from 6% to 14%. Results demonstrate superior noise suppression compared to Denoising Convolutional Neural Networks (DnCNN) and Non-Local Means (NLM), maintaining structural integrity with a Structural Similarity Index (SSIM) of 0.43 for Rician noise at σ = 6%. Segmentation performance remains stable, achieving Dice coefficients above 0.70, precision over 90%, and sensitivity between 75% to 81%, despite challenges posed by higher levels of salt-and-pepper noise. Tumor characteristics such as position and size correspond closely to ground truth, validating the effectiveness of the system in automating tumor delineation and providing reliable diagnostic assistance in neuro-oncology.

Author 1: Ngan V. T. Nguyen

Author 2: Tuan V. Huynh

Author 3: Liet V. Dang

Keywords: Magnetic Resonance Imaging (MRI); brain tumor segmentation; image denoising; Wavelet Packet Transforms (WPT); Linear Minimum Mean Square Error (LMMSE); fast c-means clustering

Paper 82: Reinventing Alzheimer’s Disease Diagnosis: A Federated Learning Approach with Cross-Validation on Multi-Datasets via the Flower Framework

Abstract: Alzheimer’s disease (AD) diagnosis using MRI is hindered by data-sharing restrictions. This study investigates whether federated learning (FL) can achieve high diagnostic accuracy while preserving data confidentiality. We propose an FL pipeline, utilizing EfficientNet-B3 and implemented via the Flower framework, incorporating advanced MRI segmentation (the Segment Anything Model, SAM) to isolate brain regions. The model is trained on a large ADNI MRI dataset and cross-validated on an independent OASIS dataset to evaluate generalization. Results show that our approach achieves high accuracy on ADNI (approximately 96%) and maintains strong performance on OASIS (around 85%), demonstrating robust generalization across datasets. The FL model attained high sensitivity and specificity in distinguishing AD, mild cognitive impairment, and healthy controls, validating the effectiveness of FL for AD MRI analysis. Importantly, this approach enables multi-center collaboration without sharing raw patient data. Our findings indicate that FL-trained models can be deployed across clinical sites, increasing the accessibility of advanced diagnostic tools. This work highlights the potential of FL in neuroimaging and paves the way for extension to other imaging modalities and neurodegenerative diseases.

Author 1: Charmarke Moussa Abdi

Author 2: Fatima-Ezzahraa Ben-Bouazza

Author 3: Ali Yahyaouy

Keywords: Federated learning; alzheimer’s disease; MRI; flower framework; data confidentiality; artificial intelligence; EfficientNet-B3; Segment Anything Model (SAM); medical image analysis; deep learning

Paper 83: Adaptive Observer-Based Sliding Mode Secure Control for Nonlinear Descriptor Systems Against Deception Attacks

Abstract: This paper delves into an advanced control scheme that combines the sliding mode control (SMC) strategy with a meta-heuristic method to examine the issue of security control for non-linear systems that are vulnerable to deception attacks on their sensors and actuators. The proposed approach focuses on the development of a secure SMC law for nonlinear descriptor systems described by TS fuzzy models. A fuzzy observer is designed to accurately estimate the states that may affected by unpredictable sensor attacks, and an adaptive SMC controller is synthesized based on the estimated information to drive the observer’s state trajectories towards the sliding surface and then maintaining the sliding motion thereafter. Afterward, sufficient conditions are established to ensure the admissibility of the closed-loop system. Then, the secretary bird optimization algorithm (SBOA), is explored for tackling an optimization problem with non-convex and nonlinear constraints as is defined to enhance the system’s performance under threats. Ultimately, a simulation study through a practical example is performed to showcase the effectiveness of the proposed control scheme in maintaining system performance, even in the presence of attacks.

Author 1: M. Kchaou

Author 2: L Ladhar

Author 3: M Omri

Author 4: R. Abbassi

Author 5: H. Jerbi

Keywords: Descriptor systems; TS fuzzy models; fuzzy observer; deception attacks; adaptive sliding mode; SBOA

Paper 84: MICRAST: Micro-Forecasting Approach for Cloud User Consumption Pattern Based on RNN

Abstract: One vital key for effective management of cloud resources is the ability to predict their users’ consumption’s patterns in granular level. It can provide more insightful analysis to guide these users towards more resource-effective habits. Such prediction requires pre-processing the users’ traces from these cloud resources for granular prediction (micro-prediction). However, the methodology followed by many forecasting based cloud studies was designed to deal with these traces as over-all trends (macro-prediction). We propose a (MICRAST) that generates segments of granular patterns. Then, it carries out parallel tasks of pre-processing and training that lead to separate trained network for each of these segments. To select a model for our approach, we compared methods from two forecasting categories: statistical and artificial neural network (ANN)-based. The results lead us to recurrent neural networks (RNN). We evaluated the MICRAST through a comparison with related work methodologies (macro-prediction approach) for both univariate and multi-variate forecasting. Then, we measured its confidence for forecasting up to 20% of the training time steps. The results showed that our approach can forecast the preferences of each cloud user with a confidence level of between (95% to 98%) surpassing related works by more than 70%.

Author 1: Shallaw Mohammed Ali

Author 2: Gabor Kecskemeti

Keywords: Micro-forecasting; cloud workload; data processing; macro-forecasting; data mining

Paper 85: Efficient Processing and Intelligent Diagnosis Algorithm for Internet of Things Medical Data Based on Deep Learning

Abstract: Electronic Medical Record (EMR) is a commonly used tool in medical diagnosis, which has static recording, difficulty in combining and analyzing different forms of data, and insufficient diagnostic efficiency and accuracy. This article proposes a CNN (Convolutional Neural Network)-LSTM (Long Short-Term Memory) algorithm for efficient processing and intelligent diagnosis of Internet of Things (IoT) medical data. The Word2Vec model is applied to clinical text data and its ability is utilized to capture semantic relationships between words. Medical image data is feature extracted using CNN, while physiological signal data is dynamically processed using LSTM to identify trends and anomalies in the data. An attention mechanism is applied to dynamically adjust the model’s attention weights for different types of data. By analyzing the samples of health, cardiovascular disease, diabetes, chronic obstructive pulmonary disease, hypertension, and chronic kidney disease, the CNN-LSTM in this article can accurately classify a variety of diseases, and the accuracy rate of healthy individuals has reached 97.8%. By combining CNN-LSTM with multimodal data, the accuracy and efficiency of medical diagnosis have been effectively improved.

Author 1: Wang Liyun

Keywords: Intelligent diagnosis; Internet of Things medical; electronic medical records; long short-term memory; convolutional neural network

Paper 86: Graph Neural Network Output for Dataset Duplication Detection on Analog Integrated Circuit Recognition System

Abstract: In the need for artificial intelligence application on the analog circuit design automation, larger and larger datasets containing analog and digital circuit pieces are required to support the analog circuit recognition systems. Since analog circuits with almost similar designs could produce completely different outputs, in case of poor netlist to graph abstraction, larger netlist input circuits could generate larger graph dataset duplications, leading to poor performance of the circuit recognition. In this study, a technique to detect graph dataset duplication on big data applications is introduced by utilizing the output vector representation (OVR) of the untrained Graph Neural Network (GNN). By calculating the multi-dimensional OVR output data into 2-dimentional (2D) representation, even the random weighted untrained GNN outputs are observed to be capable of distinguishing between each graph data inputs, generating different output for different graph input while providing identical output for the same duplicated graph data, and allowing the dataset’s duplication detection. The 2D representation is also capable of visualizing the overall datasets, giving a simple overview of the relation of the data within the same and different classes. From the simulation result, despite being affected by the floating-point calculation accuracy and consistency deficiency, the F1 score using floating-point identical comparisons are observed with an average of 96.92% and 93.70% when using CPU and GPU calculations, respectively, while the floating-point rounding calculation is applied. The duplication detection using floating point range comparison is the future work, combined with the study of the 2D GNN output behavior under the ongoing training process.

Author 1: Arif Abdul Mannan

Author 2: Koichi Tanno

Keywords: Big data; graph neural network; artificial intelligence; analog circuit design

Paper 87: Advanced Image Recognition Techniques for Crop Pest Detection Using Modified YOLO-v3

Abstract: Accurate and efficient detection of agricultural pests is crucial for crop protection and pest control. This study addresses the limitations of traditional pest detection methods, such as weak detection capabilities and high computational demands, by proposing an improved image recognition system based on the YOLO-v3 algorithm. The research focuses on enhancing pest detection accuracy through deep learning techniques, specifically by modifying the YOLO-v3 model with the ISODATA clustering algorithm, DenseBlock enhancements, and the ELU activation function. A dataset of 13,000 images representing six common crop pests was created and expanded using various image augmentation techniques. The modified YOLO-v3 model was trained and evaluated on this dataset, achieving a higher mean Average Precision (mAP) of 89.7% and faster recognition speed compared to Faster-RCNN, SSD-300, and the original YOLO-v3 model. Finally, the improved model demonstrated a recognition speed of 27 frames per second (fps), significantly outperforming other detection models in both accuracy and speed. The proposed method offers a superior solution for real-time pest detection in agricultural settings, combining high accuracy with computational efficiency. Future work will explore the application of optimization algorithms to further enhance the robustness and generalizability of the system across diverse pest detection scenarios.

Author 1: Dechao Guo

Author 2: Hao Zhang

Keywords: Feature detection algorithm; YOLO-v3 network; image recognition technology; crop pest detection applications

Paper 88: CodifiedCant: Enhancing Legal Document Accessibility Using NLP and Longformer for Secure and Efficient Compliance

Abstract: CodifiedCant is a new idea that employs Natural Language Processing to simplify company guidelines and legal documents. Legal texts are extensive, complicated and hard for non-experts to understand. To tackle the above problem, this research incorporates the Longformer model because it functions as a transformer-based deep learning system designed to work effectively with extensive legal documents. Longformer enables the system to handle extensive documents by keeping better track of context, which results in transforming complex legal text into easily readable formats. To enhance the search and retrieval speed, this research investigates the nuances of transforming unstructured data, like tabular data from PDFs, to vectors. This revolution supports quicker, cognisant semantic routing inside the document. Further, it assists in data arrangement and detection across massive sources of legitimate and business information. Data security is also a major priority for the platform, which utilizes network encryption to protect data and privacy. CodifiedCant is a scalable, secure and intelligent solution for better employee access to legal news, greater company transparency and reinforces better compliance in the organization. Table extraction and document simplification performance of the model are validated on Cornell LII and Kaggle evaluation datasets, respectively. CodifiedCant associates the variance relating to legitimate terminology and user knowledge.

Author 1: Jayapradha J

Author 2: Su-Cheng Haw

Author 3: Naveen Palanichamy

Author 4: Nilanjana Bhattacharya

Author 5: Aayushi Agarwal

Author 6: Senthil Kumar T

Keywords: Natural language processing; transformer-based deep learning system; long former; semantic routing; network encryption; legal document; unstructured data; and data security

Paper 89: Multi-Dimensional Digital Media Sentiment Visualization Intelligent Analysis System Based on Machine Learning Algorithm

Abstract: This study builds a multi-dimensional sentiment analysis system to solve the problem of sentiment prediction of text and image data in the Weibo platform. By combining CNN (Convolutional Neural Network), BiLSTM (Bidirectional Long Short-Term Memory) and Attention mechanism (AM), the accuracy of sentiment classification is improved, which helps to better understand and analyze user sentiment expressions in social media. This study uses crawler tools to collect text and image data of 1,000 users on the Weibo platform from January to December 2021 to ensure the diversity and representativeness of the data; the text data is segmented, stop words are removed, and the text is converted into vectors; at the same time, the ResNet-50 pre-trained model is used to extract the deep features of the image, CNN is used to process the image data, and BiLSTM captures the contextual information in the text data. Finally, the AM is used to enhance the model's attention to emotional expression. Experimental results show that the proposed Word2Vec (Word to Vector) model performs outstandingly in the accuracy of sentiment classification. The accuracy of the CNN-BiLSTM-Attention model in positive, neutral and negative classification tasks is 97.5 per cent, 95.4 per cent and 91.6 per cent, respectively, which are significantly better than the performance of the CNN and BiLSTM models, especially in the evaluation indicators such as accuracy and macro F1. This study proposes a multimodal sentiment analysis system based on CNN-BiLSTM-Attention, which significantly and effectively improves the accuracy of social media sentiment classification. The system can effectively process complex sentiment categories and multimodal data, and has broad application prospects, especially in the fields of social media sentiment analysis and public opinion monitoring.

Author 1: Mengwei Leia

Author 2: Qiong Chen

Keywords: Digital media; sentiment analysis; intelligent systems; multimodal data

Paper 90: Emotion-Aware EEG Analysis for Alzheimer’s Disease Detection Using Boosting and Deep Learning

Abstract: Alzheimer’s disease (AD) is a leading cause of dementia, yet its diagnosis remains challenging. EEG provides a noninvasive and cost-effective method for monitoring brain activity, which may reflect both cognitive decline and altered emotional states. In this study, an EEG-based pipeline was developed to classify AD using two approaches: an ensemble of boosting classifiers based on extracted features, and a deep convolutional neural network (CNN) applied to raw signals. A publicly available dataset was processed to extract time, frequency, and complexity features, with emotional brain dynamics implicitly reflected in the signals and considered during analysis. Five ensemble models (including CatBoost, LightGBM, and XGBoost) were optimized using Bayesian search. The CNN was trained separately and evaluated under cross-validation schemes. A balanced accuracy of 78.96% was achieved for AD detection using XGBoost, while the CNN reached 70.92% for Frontotemporal dementia. The study demonstrates that combining machine learning with EEG produces generalizable models for dementia detection and suggests that accounting for emotion-related variability may enhance diagnostic results.

Author 1: Shynara Ayanbek

Author 2: Abzal Issayev

Author 3: Amandyk Kartbayev

Keywords: Alzheimer’s disease; feature extraction; machine learning; CNN; boosting algorithms; deep learning

Paper 91: The Impact of Federated Learning on Distributed Remote Sensing Archives

Abstract: When it comes to Machine Learning in remote sensing, one of the main obstacles researchers face is the large scale of datasets. Just the size of freely available Earth observation data presents a challenge for personal computers. A variety of missions, such as Sentinel-1, -2, and -3, have collectively gathered several petabytes of data. Given the size of these datasets, they are stored and processed across multiple platforms (often referred to as clients), which implies that decentralized Machine Learning must be applied. Federated Learning is one such decentralized learning approach, originally introduced by Google and adopted in their Android ecosystem. Since its release, the original Federated Learning technique has been fine-tuned and further developed. The scope of this project is to apply multiple Federated Learning models on remote sensing datasets and understand their implications considering different data splits across clients.

Author 1: Pratik Surendrakumar Patel

Author 2: Vijay Govindarajan

Keywords: Machine learning; federated learning; deep learning model

Paper 92: An Event-B Capability-Centric Model for Cloud Service Discovery

Abstract: Cloud computing has become increasingly adopted due to its ability to provide on-demand access to computing resources. However, the proliferation of cloud service offerings has introduced significant challenges in service discovery. Existing cloud service discovery approaches are often evaluated solely through simulation or experimentation and typically rely on unstructured service descriptions, which limits their precision and scalability. In this work, we address these limitations by proposing a formally verified architecture for capability-centric cloud service discovery, grounded in the Event-B method. The architecture is built upon a capability-centric service description model that captures service semantics through property-value representations. A core element of this model is the formally verified variantOf relation, which defines specialization among services. We prove that variantOf satisfies the properties of a partial order, enabling services to be structured as a Directed Acyclic Graph (DAG) and thus supporting hierarchical and scalable discovery. We formally verify the consistency of our model across multiple refinement levels. All proof obligations generated by the Rodin platform were successfully discharged. A scenario-based validation further confirms the correctness of dynamic operations within the system.

Author 1: Aicha Sid’Elmostaphe

Author 2: J Paul Gibson

Author 3: Imen Jerbi

Author 4: Walid Gaaloul

Author 5: Mohamedade Farouk Nanne

Keywords: Formal verification; cloud service discovery; capability modelling

Paper 93: Analyzing the Impact of Histogram-Based Image Preprocessing on Melon Leaf Abnormality Detection Using YOLOv7

Abstract: This study aims to analyze and implement image preprocessing techniques to improve the performance of melon leaf abnormality detection using the YOLOv7 algorithm. A total of 521 abnormal melon leaf images were processed using augmentation and three preprocessing methods: Averaging Histogram Equalization (AVGHEQ), Brightness Preserving Dynamic Histogram Equalization (BPDFHE), and Contrast Limited Adaptive Histogram Equalization (CLAHE), then compared with the original dataset. Modeling was conducted in three stages: initial training with an 80:20 split and default YOLOv7 augmentation; hyperparameter tuning via cross-validation using a 90:10 split without augmentation; and final training using the best parameters with augmentation reactivated. The models were evaluated using ensemble learning. Results showed mAP ranged from 58.6% to 66.3%, accuracy from 80.7% to 84.9%, and detection time from 9.8 to 20 milliseconds. Preprocessing improved mAP and detection time, though it had little effect on accuracy. The best performance was obtained with a kernel size of 3 and a learning rate of 0.001, while changes in activation function, pooling, batch size, and momentum had minimal impact. The top models, trained with maximum epochs and standard augmentation, achieved mAP of 84.12%, accuracy of 91.19%, and detection time of 4.55 milliseconds. Models using early stopping (patience = 300) reached mAP of 81.57%, accuracy of 92.23%, and detection time of 5.03 milliseconds. The best model outperformed previous works, which reported only 48.85% with Faster R-CNN, 33.16% with SSD, and 16.56% with YOLOv3. Although histogram-based preprocessing methods mainly enhanced inference speed, the overall improvements to YOLOv7 significantly boosted detection performance.

Author 1: Sahrial Ihsani Ishak

Author 2: Sri Wahjuni

Author 3: Karlisa Priandana

Keywords: Leaf abnormality; melon; image preprocessing; YOLOv7

Paper 94: Remote Monitoring and Management System for Oil and Gas Facilities with Integrated IoT and Artificial Intelligence Data Analysis

Abstract: An important trend in the development of digitalization is the expansion from digitalization to intelligence. For the intelligence of oil and gas facilities, it is to apply the multi-intelligent judgment and analysis technology to the actual production of oil and gas facilities, so as to realize the real-time data acquisition, analysis and “monitoring tube” integration of remote oil and gas facilities. As a flexible monitoring configuration software, force control provides a good development interface and a simple engineering implementation method under the trend of intelligent oil and gas facilities, which can facilitate users to realize and complete the functions of the monitoring layer. Its humanized application, such as trend curve, report output, limit alarm and other system functions, can also be more intuitive for users to monitor and process control on-site production data, and provide strong technical support for the intelligent process of oil and gas facilities. Based on this, this paper studies the design of remote monitoring and management system of oil and gas station by integrating the Internet of Things, artificial intelligence and big data analysis technology.

Author 1: Shu Haowen

Author 2: Zhang Bin

Author 3: Gao Shiyu

Author 4: Gu Li

Author 5: Jia Yanjie

Keywords: Internet of Things; artificial intelligence; data analytics; remote monitoring; management system

Paper 95: Innovative Design Algorithm of Huizhou Bamboo Weaving Patterns Based on Deep Learning

Abstract: In the field of innovative design of Huizhou bamboo weaving patterns, traditional deep learning algorithms cannot fully capture the fine structure and subtle changes of patterns, resulting in distorted or blurred results, and require a lot of computing resources and time during the training process. This paper constructs an improved ViT (Vision Transformer) model to collect diverse Huizhou bamboo weaving pattern data covering different styles and forms. In the data enhancement stage, common enhancement techniques such as rotation, scaling, flipping, and color perturbation are used to increase the diversity of training data. Based on the traditional ViT model, a local self-attention mechanism is applied to replace the traditional global self-attention mechanism. Mixed precision training and distributed training strategies are used to effectively accelerate the training process while maintaining high accuracy. The model automatically generates innovative designs by learning the style and structural characteristics of Huizhou bamboo weaving patterns, and adds a detail repair module in the generation process to enhance the detail expression of the pattern. The experimental results show that the improved ViT model tends to 0.95 after 50 training rounds, indicating that it performs well in detail preservation and structural similarity; with a sample data volume of 5000, the training time of the improved ViT model is 47.4 seconds, and the GPU memory usage is 37.1GB, providing higher computing efficiency. The experimental results prove the effectiveness of this paper’s research on the innovative design algorithm of Huizhou bamboo weaving patterns.

Author 1: Jinjin Rong

Author 2: Xin Fang

Keywords: Deep learning; Huizhou bamboo weaving; bamboo weaving pattern; vision transformer; local self-attention mechanism