Towards Robust Federated Learning: A Centroid-Based Approach to Jointly Mitigate Noisy Labels and Non-IID Data

Abstract

In federated learning (FL), client data often suffer from the challenges of data distribution imbalance such as Non-IID and noisy labels. Crucially, these two issues are highly coupled and mutually exacerbating: Non-IID data complicates the identification of noisy labels, while noisy labels severely amplify local model drift. The compounding effect leads to severe training instability and degrades model convergence. To tackle these issues, we propose FedCRT, a Federated Centroid-based Robust Training framework designed to jointly mitigate both issues. At the client level, we introduce a centroid-based noisy label correction mechanism and class-conditional feature centroids. To prevent confirmation bias during this process, we devise a dual-confidence strategy for cautious and accurate label correction, accompanied by a robust loss function to maintain consistent optimization directions. At the server level, we develop a two-stage quality-aware aggregation strategy. It not only mitigates Non-IID-induced client drift but also dynamically assigns greater aggregation weights to clients with higher estimated clean data proportions. Extensive experiments have been conducted on benchmark datasets. The experiment results demonstrate that our method performs well in noisy label detection and correction on different clients under highly heterogeneous label noise scenarios which prove its effectiveness and robustness.