Optimized federated-learning-based techniques for IoT in edge computing

Abstract

The Internet of Things (IoT) generates vast amounts of data from billions of interconnected devices, creating unprecedented opportunities for machine learning applications. However, traditional centralized learning approaches face significant limitations in IoT environments, including privacy concerns, bandwidth constraints, and data sovereignty issues. Federated Learning (FL) emerges as a promising solution that enables collaborative model training across distributed edge devices while keeping data locally stored, thus preserving privacy and reducing communication overhead. This work addresses four critical challenges in federated learning for IoT systems: (1) Communication Efficiency - reducing bandwidth requirements through optimized synchronization strategies in resource-constrained networks; (2) Data Heterogeneity and Multimodality - handling non-identically distributed (non-i.i.d.) data across devices and managing missing modalities in multimodal IoT environments; (3) System Heterogeneity - addressing computational resource variations and unpredictable client participation patterns; and (4) Lifelong Learning - managing concept drift in evolving data distributions and mitigating the straggler problem in dynamic IoT deployments. The research focuses on optimizing FL strategies to maintain or improve performance under realistic constraints, moving beyond idealized assumptions to address the practical challenges of implementing federated learning in large-scale, heterogeneous IoT networks. By tackling these fundamental issues, this work aims to enable more efficient and robust federated learning systems that can effectively leverage the collective intelligence of distributed IoT devices while respecting privacy and resource constraints.