Managing Heterogeneity in Client Selection for IoT-Based Federated Edge Learning

The integration of the Internet of Things (IoT) with edge computing is driving a paradigm shift in how intelligent services and applications are deployed, particularly in the era of 6G networks. IoT devices at the edge generate vast amounts of data, necessitating the development of Machine Learning (ML) models tailored to specific applications such as autonomous vehicles, smart healthcare, and industrial automation. Federated Edge learning (FEEL), a privacy-preserving approach to distributed ML, has emerged as a key enabler for such use cases by allowing IoT devices to collaboratively train models without sharing raw data. However, the heterogeneity of IoT devices in terms of data distributions, computational resources, and energy constraints poses significant challenges to efficient model training in FEEL. In this context, addressing the client selection problem becomes critical, as random sampling often leads to suboptimal utilization of device updates, resulting in slower convergence, degraded model accuracy, and reduced fairness. Given the NP-hard nature of the client selection problem, we propose a heuristic client selection scheme tailored for FEEL in IoT-edge environments. Our approach incorporates criteria such as weight divergence, training deadlines, and reliability measure to ensure the overall efficiency of the learning process. Extensive experimental evaluations demonstrate that our scheme consistently outperforms baseline methods, including random selection, speed-based selection, and weight divergence-based selection, achieving higher global accuracy, lower round training time, and significantly reduced dropout ratios.