Continual Local Updates for Federated Learning with Enhanced Robustness to Link Noise

Communication errors caused by noisy links can negatively impact the accuracy of federated learning (FL) algorithms. To address this issue, we introduce an FL algorithm that is robust to communication errors while concurrently reducing the communication load on clients. To formulate the proposed algorithm, we consider a weighted least-squares regression problem as a motivating example. We recast this problem as a distributed optimization problem over a federated network, which employs random scheduling to enhance communication efficiency, and solve the reformulated problem via the alternating direction method of multipliers. Unlike conventional FL approaches employing random scheduling, the proposed algorithm grants the clients the ability to continually update their local model estimates even when they are not selected by the server to participate in FL. This allows for more frequent and ongoing client involvement, resulting in performance improvement and enhanced robustness to communication errors compared to when the local updates are only performed when the respective clients are selected by the server. We demonstrate the effectiveness and performance gains of the proposed algorithm through simulations.