Resource-Aware Split Federated Learning for Edge Intelligence

This research investigates Federated Learning (FL) systems, wherein multiple edge devices cooperate to train a collective machine learning model using their locally distributed data. The focus is on addressing energy consumption challenges in battery-constrained devices and mitigating the negative impact of intensive on-device computations during the training phase. In the widespread adoption of FL, variations in clients' computational capabilities and battery levels lead to system stragglers/dropouts, causing a decline in training quality. To enhance FL's energy efficiency, we propose EAFL+, a pioneering cloud-edge-terminal collaborative approach. EAFL+ introduces a novel architectural design aimed at achieving power-aware FL training by capitalizing on resource diversity and computation offloading. It facilitates the efficient selection of an approximately-optimal offloading target from Cloud-tier, Edge-tier, and Terminal-tier resources, optimizing the cost-quality tradeoff for participating client devices in the FL process. The presented algorithm minimizes the dropouts during training, enhancing participation rates and amplifying clients' contributions, resulting in better accuracy and convergence. Through experiments conducted on FL datasets and traces within a simulated FL environment, we find EAFL+ eradicates client drop-outs and enhances accuracy by up to 24% compared to state-of-the-art methods.