Anonymous and Byzantine-Robust Federated Learning With Secure and Efficient Aggregation

Federated learning (FL) serves as a distributed machine learning framework that addresses the challenges of data silos while preserving data privacy. Specifically, FL enables multiple participants to collaboratively train a global model by sharing local updates without exposing their raw local data. Although FL achieves physical data isolation through local update sharing mechanisms, it still faces emerging security threats. On the one hand, adversaries may reconstruct sensitive data features or infer client attributes by analyzing local updates. On the other hand, clients might upload malicious updates to disrupt global model aggregation, causing performance degradation. To solve these issues, we propose an anonymous and Byzantine-robust FL scheme with secure and efficient aggregation. First, we propose a single-masking protocol that not only preserves data privacy but also enhances aggregation efficiency. Second, we eliminate client message metadata, such as source IP addresses and timestamps, through secure shuffling, achieving client anonymity in conjunction with the single-masking protocol. Additionally, we implement a baffle mechanism to resist the impact of malicious updates on the global model, thereby ensuring Byzantine robustness. Security analysis demonstrates that our scheme simultaneously preserves data privacy and identity anonymity. Experimental results show that our scheme can effectively resist poisoning attacks, even if 50% of the fog nodes are contaminated by malicious clients. Moreover, the aggregation efficiency of the proposed scheme is improved by over 20%.