Solving Client Dropout in Federated Learning via Client Similarity Discovery and Gradient Supplementation Mechanism

Abstract

In the realm of practical applications of federated learning, an issue arises wherein the performance suffers due to passive client disconnections during the federated training process, caused by factors such as resource limitations or network disruptions. This paper introduces a More Precise Similarity Discovery and Gradient Supplementation(MPSDGS) algorithm, which tackles the problem of passive client dropout in federated learning by employing precise clustering techniques to identify similar clients. It further leverages the gradients of clients whose data distribution closely aligns with the disconnected clients, effectively supplementing the disconnected client gradients. The algorithm’s efficacy is verified through experimental evaluations conducted on real-world datasets, namely MNIST, CIFAR10, and CIFAR100. The experimental findings reveal that, under the same non-independent and identically distributed data partitioning approach for MNIST and CIFAR10 datasets, MPSDGS achieves notable accuracy enhancements. Specifically, at disconnection rates of 0.3, 0.5, and 0.7, the MPSDGS algorithm improves the accuracy of the MNIST dataset by 1.33%, 1.49%, and 1.35%, respectively. Similarly, for the CIFAR10 dataset, the algorithm enhances accuracy by 1.09%, 1.25%, and 1.6%, respectively, at the aforementioned disconnection rates. Remarkably, MPSDGS exhibits comparable excellence in performance on the CIFAR100 dataset.