Uplink Time Constrained Federated Learning over Wireless Networks
Ji Ho Choi, Dong In Kim · 2023
Federated learning (FL) is an emerging distributed learning paradigm that collaboratively trains a shared model while preserving their data privacy. However, clients participating in the FL process have heterogeneous computation power and communication resources. Therefore, the central server needs to wait for the slowest client finishes uploading its local model update, which is known as straggler effect. In this paper, we propose a novel FL method with uplink time constraint set by the server to mitigate the straggler effect over wireless networks. First, we study the impact of uplink time constraint on the number of clients with successful upload in Rayleigh fading channel. Since acquiring perfect channel state information (CSI) is challenging in practical FL applications, the server exploits channel statistics instead. We then formulate an optimization problem for the error convergence speed with respect to uplink time constraint. Finally, we confirm that the proposed scheme, namely FL with Uplink Time Constraint (FedUTC), achieves faster convergence speed than the baseline FL method.