ADaRES: Adaptive Data Reallocation and Epoch Scheduling in Federated Learning Under Network and Resource Constraints

Federated Learning (FL) efficacy in practical edge and IoT environments is often hindered by client unavailability and non-IID data. This paper addresses the essential problem of client unavailability, in which data from entire classes may become inaccessible. We present ADaRES, a framework for Adaptive Data Reallocation and Epoch Scheduling. We provide an overview of the ADaRES system, which detects clients with limited resource that are reachable via network, redistributes their data to stable peers, and adjusts local epochs to evenly the new computational load. We assess ADaRES within a challenging simulation setup based on Perlin noise, which simulates realistic, time dependent client availability. Our experiments reveal that ADaRES achieves a final accuracy (76.9%) comparable to that of an ideal system with constant availability, while a conventional FL model in the same unstable environment fails to converge (49.1%). We conclude that ADaRES serves as a computationally-efficient and resilient solution to the challenge of client unavailability in FL.