FedCGSU: Client Grouping Based on Similar Uncertainty for Non-IID Federated Learning

Federated Learning (FL) is an approach that allows nodes with limited resources to collaborate without the need to share their data. However, the computational performance of local devices, the the non-independent and identically distributed (Non-IID) nature of data, and limited communication resources inevitably reduce the model's convergence speed and accuracy. It is urgent to obtain FL models with superior performance under Non-IID data. In this paper, we propose FedCGSU, a federated learning method for client grouping based on similar uncertainty in local client distributions. FedCGSU achieves grouping by leveraging the similarity in local distributions, which reduces the impact of weight divergence among clients with different distributions. It then combines loss values and local data volumes in the aggregation method, which reduces the impact of computationally weak devices on the global convergence speed. Experiments using three public datasets are conducted, demonstrating that the FedCGSU approach exhibits outstanding performance in improving accuracy and accelerating convergence speed.