Harnessing the Power of Federated Learning: A Systematic Review of Light Weight Deep Learning Protocols

With rapid proliferation in using smart devices, real time efficient sentiment analysis has gained considerable popularity. These devices generate variety of data. However, for resource con-strained devices to perform sentiment analysis over multimodal data using conventional modals that are computationally complex and resource hungry, is challenging. This challenge may be addressed using a light weight but efficient modal specifically focused on sentiment analysis for contrained devices. in the literature, there are several modals that claims to be light weight however, the real sense and logic to determine if the modal may be termed as lightweight still requires further research. This paper reviews approaches to federated learning for multimodal sentiment analysis. Federated learning enables decen-tralized training without sharing data. Considering the review need to balance privacy concerns, performance, and resource usage, the review evaluates existing approaches to enhance accuracy in sentiment classification. The review identifies strengths and limitations in handling multimodal data. The search focused on studies in databases like IEEE Xplore and Scopus. Studies published in peer-reviewed journals over the past five years were included. The review covers 45 studies, mostly experimental, with some theoretical models. Key results show lightweight protocols improve efficiency and privacy in federated learning. They reduce computational demands while handling text, image, and audio data. There is a growing focus on resource-constrained devices in research. Trade-offs between model complexity and speed are commonly explored. The review addresses how these protocols balance accuracy and computational cost.

@article{Shinwari2025,
  title     = {Harnessing the Power of Federated Learning: A Systematic Review of Light Weight Deep Learning Protocols},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {1},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Haseeb Khan Shinwari and Riaz Ul Amin},
  doi       = {10.14569/IJACSA.2025.0160107},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160107}
}