Optimized Threshold Selection for Client Participation in Iot-Based Federated Learning

Abstract

Implementing Federated Learning (FL) in IoT environments presents notable challenges, primarily due to variations in both clients' local model performance and the reliability of their communication links, often indicated by received signal strength (RSS). In this study, we introduce a deterministic and lightweight selection mechanism that filters participants based on predefined thresholds for local model accuracy (LMA) and RSS. The thresholds are analytically optimized to maximize the probability of selecting exactly <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m$</tex> suitable clients. Theoretical analysis is complemented by parameter-specific insights to better illustrate the system's behavior. Comprehensive experiments using the MNIST dataset demonstrate that our approach consistently surpasses strategies that rely solely on LMA or RSS, validating its effectiveness for FL in resource-limited IoT contexts.