POCS: phased optimal client selection for federated learning

Abstract Traditional federated learning suffers from excessive communication overhead due to iterative model exchanges, long training time caused by data heterogeneity, and vulnerability to adversarial attacks due to gradient leakage risks. But if one strategically selects some specific clients for training during different training rounds, one can reduce the costs and the training time, while defending against the adversarial attacks. In this work, we propose the Phased Optimal Client Selection for Federated Learning (POCS) method, which includes customized optimal Client Selection method for each specific training phase in federated learning. We evaluate POCS’s performance with that of five state-of-the-art baselines, while the results indicate that POCS can improve the accuracy, reduce the training time and the communication costs, and show better defense against the membership inference attacks.