A Robust Byzantine-Resilient Framework for Federated Learning
Abstract / Description
Abstract
The features of Federated Learning (FL) render it highly vulnerable to Byzantine attacks, drawing significant research attention. However, the non-independently and homogeneously distributed(non-IID) nature of client data introduces two key challenges: (i) existing approaches typically rely on a single evaluation dimension—such as validation score or update similarity—which can lead to potential omissions or misclassifications, and (ii) under non-IID conditions, it is difficult to differentiate between clients with naturally unique data distributions and those exhibiting malicious behavior. To address these issues, we propose an efficient detection strategy that groups clients and employs a divide-and-conquer approach to integrate multidimensional information. By combining cross-validation with data similarity analysis, our approach aims to identify malicious behaviors more accurately, reduce false positives, and enhance the robustness of the global model as well as the resilience of FL against Byzantine attacks.