EnCAgg: Adaptive Clustering Defense for Federated Learning Against Dynamic Model Poisoning

arXiv·low signal

EnCAgg improves federated learning robustness by replacing fixed-cluster defenses with enhanced clustering aggregation that dynamically adapts to evolving poisoning strategies. Existing defenses use fixed thresholds or fixed cluster counts to separate malicious from benign gradients, but these approaches fail against adaptive attackers and discard benign gradients from heterogeneous client data. The adaptive method better preserves model utility while filtering poisoned updates.