FedSW: A Sliding Window-Based Approach for Asynchronous Federated Learning in WiFi Networks

Federated learning (FL) presents a novel paradigm for constructing global models by leveraging distributed client data while preserving privacy. Despite clients’ readiness to contribute computational resources via WiFi networks, the concurrent model uploads often trigger the competitive backoff mechanism inherent in the carrier sense multiple access with collision avoidance (CSMA/CA) protocol, which impairs the training efficiency and performance of FL. To address this challenge, this paper proposes an innovative sliding window-based asynchronous update approach for federated learning, named as FedSW. By properly configuring the sliding window size at the wireless access point (AP) of the WiFi network, FedSW orchestrates local training, model upload, aggregation, and distribution in harmony with the sliding window progress. This synchronization significantly improves training efficiency and model performance. Furthermore, the versatility of FedSW is demonstrated through its seamless integration with state-of-the-art (SOTA) algorithms. Our methodology is rigorously evaluated against FL benchmarks, showcasing its superior effectiveness. Simulation results confirm that FedSW consistently outperforms conventional benchmarks in terms of convergence, regardless of the sliding window size, while significantly reducing latency.