Robust Federated Learning Based on Chained Self-Paced Learning

Federated learning is a framework that enables distributed collaborative training while preserving data privacy, and it has great potential for enhancing diagnostics in the medical field. Due to the reluctance of participating institutions to disclose or share details of their models, they design their own network architectures. However, this heterogeneity in models makes it challenging to ensure that all client-collected data in federated learning is correctly annotated, inevitably leading to noise. Therefore, we propose a spiral self-training strategy to address the issue of internal client noise, combating corrupted label noise through reweighted aggregation updates across different time periods. Our proposed federated learning approach improves the original baseline by 5.30% under a noise rate of 0.1 and the pairflip noise scheme. Under symflip noise, it achieves a 5.87% improvement. Similarly, at a noise rate of 0.2, there is a substantial improvement. Extensive experiments demonstrate that our method significantly outperforms recent state-of-the-art methods on noisy label data.