SSFed: Statistical Significance Aggregation Algorithm in Federated Learning

Federated learning enables collaborative model training across multiple clients without sharing raw data, where the global server aggregates local models. One of the primary challenges in this setting is dealing with non-i.i.d data, which can lead to biased aggregations, as well as the overhead of frequent communication between clients and the server. Our approach improves state-of-art aggregation by adding statistical significance testing. This step assigns greater weight to client updates with higher statistical impact. Only statistically significant updates are included in the global model. The process begins with each client training a local model on its dataset. Clients then send these trained parameters to the server. At the global server, statistical significance testing is applied by calculating z-scores for each parameter. Updates with z-scores below a set threshold are included, with each update weighted based on its significance.SSFed achieves a final accuracy of 88.71% in just 20 rounds, outperforming baseline algorithms and resulting in an average improvement of 25% over traditional federated learning methods. This demonstrates faster convergence and stronger performance, especially under highly non-i.i.d client data distributions. Our SSFed implementation is available on GitHub.