Disentangling Client Contributions: Improving Federated Learning Accuracy in the Presence of Heterogeneous Data

Federated Learning (FL) is a promising paradigm that leverages distributed data sources to train machine learning models, thereby offering significant privacy advantages. However, the inherent statistical heterogeneity among clients, characterized by distinct data distributions and varying model performance, poses a substantial challenge. Such heterogeneity adversely affects both the convergence and overall performance of the global model within the FL framework. To address this issue, we introduce a novel FL algorithm, FedVa, which considers both the local data volume and model accuracy to determine client contributions, assigning respective weights. This consideration facilitates greater engagement from key clients during the aggregation process, thereby enhancing their influence on the global model refinement. Experimental results indicate that FedVa surpasses prevailing methods in managing statistical heterogeneity and enhancing global model accuracy, without incurring additional communication costs. Specifically, FedVa achieves a convergence speed 3.9 times faster than FedAvg on the MNIST dataset, and on the CIFAR-10 dataset, it results in a 2.3% increase in final model accuracy compared to FedAvg. Our code is publicly available at https://github.com/ChunmingLiu23/FedVa.