Federated Dynamic Client Selection Based on Comprehensive Performance Evaluation

Federated Learning, as a distributed machine learning approach, enables the full exploitation of data value while protecting data privacy. However, traditional Federated Learning methods are significantly impacted in terms of model accuracy and convergence speed when dealing with Non-Independent and Identically Distributed (Non-IID) data. This paper introduces a Federated Dynamic Client Selection method based on Comprehensive Performance Evaluation (FedDCP), which dynamically adjusts the probability of client participation in training by integrating the clients' computational capabilities, historical performance, and current performance. This approach optimizes model accuracy, convergence speed, and runtime. Experimental results indicate that the FedDCP method outperforms existing Federated Averaging optimization methods in terms of model accuracy and convergence speed on the MNIST, Fashion-MNIST, and Cifar-10 datasets, with a similar runtime, demonstrating its effectiveness.