CSMAAFL: Client Scheduling and Model Aggregation in Asynchronous Federated Learning
Xiang Ma, Qun Wang, Haijian Sun, Rose Qingyang Hu, Yi Qian · 2024
Asynchronous federated learning aims to solve the straggler problem in an environment with heterogeneity, where certain clients may possess limited computational capacities, potentially leading to model aggregation delay. The core concept behind asynchronous federated learning is to empower the server to aggregate the model as soon as it receives an update from any client without waiting for updates from multiple clients or adhering to a predetermined waiting time, which is typical in synchronous mode. Because of the asynchronous setting, a potential concern is the emergence of a stale model issue, wherein slow clients might employ an outdated local model for their data training. Consequently, when these locally trained models are uploaded to the server, they may impede the convergence of the global training. Therefore, effective model aggregation strategies play a significant role in updating the global model. Besides, client scheduling is critical when heterogeneous clients with diversified computing capacities participate in the federated learning process. This work first investigates the impact of the convergence of asynchronous federated learning mode when adopting the aggregation coefficient in synchronous mode. Effective aggregation solutions that can achieve the same convergence result as in the synchronous mode are proposed, followed by an improved aggregation method with client scheduling. The simulation results in various cases demonstrate that the proposed algorithm converges with a similar level of accuracy as the classical synchronous federated learning algorithm but effectively accelerates the learning process, especially in its early stage.