FedCRA: Client Risk-aware Aggregation via Update Reconstruction for IoT-based Federated Learning

Abstract

Federated Learning (FL) enables distributed model training without sharing raw data, making it particularly suitable for Internet of Things (IoT) environments. However, FL remains vulnerable to poisoning attacks, where malicious clients manipulate local updates to corrupt global model optimization. Existing defenses mainly focus on updates within a single round and may confuse heterogeneous benign updates with malicious ones. In this paper, we propose FedCRA, a client risk-aware aggregation framework that exploits historical trajectories and trusted reference updates to reconstruct updates. Furthermore, the risk-aware aggregation jointly considers reconstruction gap, reference gap, and historical trajectory information to adaptively suppress suspicious updates. Extensive experiments under multiple poisoning attacks and heterogeneous data distributions demonstrate that FedCRA achieves stronger robustness than several representative aggregation methods while maintaining good scalability in large-scale FL.