Robust Federated Learning via Loss-Based Client Clustering

A novel strategy for resilient federated learning (FL) against Byzantine attacks incorporates loss-based client clustering. This technique operates under the assumption of a reliable server equipped with a small auxiliary dataset and necessitates only two trustworthy participants (the server and one client) to be effective, without needing to know the count of malicious clients in advance. Theoretical evaluations indicate that it achieves bounded optimality gaps, even amidst severe attacks. Experimental results reveal that it significantly surpasses traditional and robust FL benchmarks—Mean, Trimmed Mean, Median, Krum, Multi-Krum—across various attack methods, including label flipping. The research can be accessed on arXiv (2508.12672).