RARE-FL: Resilient Accelerated and Risk-Aware Edge Federated Learning in Scarce Data Scenario

Federated learning as a Service (FLaaS) is promoted as a privacy-preserving collaborative machine learning, which is challenged by channel impairments and the scarcity of trustworthy devices in wireless networks. Assuming a trustworthy metric (TM) reflecting the reported local models accuracy, we propose a resilient, accelerated, and risk-aware edge FL (RARE-FL) that utilizes the TM scores to provide fast and trustworthy FLaaS without relying on validation datasets. The proposed RARE-FL is applied to a communication harsh 6G non-terrestrial network, where mutually interfering unmanned aerial vehicles (UAV) provide FLaaS to spatially distributed edge devices. Our numerical results on MNIST and CIFAR-10 datasets affirm the efficacy of the proposed RARE-FL, showcasing its enhanced performance and reliability over existing methodologies.