A Buffered Semi-Asynchronous Mechanism with MAB for Efficient Federated Learning
Chenyu Wang, Qiong Wu, Qian Ma, Xu Chen · 2022
Federated learning can leverage abundant data generated by scattered client devices without impairing their data privacy. However, heterogeneous properties of devices may cause a long single round duration in synchronous federated learning and stale gradient of clients in asynchronous federated learning. To solve these issues, we propose a buffered semi-asynchronous federated learning with efficient client selection strategy, named BSACS-FL. Specifically, we introduce a buffer mechanism which is used to store models that have not yet been aggregated on client side to improve clients computing resource utilization. In addition, we adopt Multi-Armed Bandit (MAB) method to select a proper set of clients to upload new local models based on the waiting time of local model in the buffer and the learned training capability of clients. The novel BSACS-FL algorithm boosts time efficiency and improves client computing resource utilization, which achieves more effective model aggregations, leading to a faster convergence with high model accuracy. Extensive experiments based on both image classification and human activity recognition tasks demonstrate the superior performance of BSACS-FL, e.g., improving the model accuracy by 19% compared to the FedAvg.