A Privacy-Preserving Federated Learning Framework with Byzantine Robustness

Federated learning, an innovative privacy-preserving distributed machine learning framework, allows participants from various parties to co-construct models without directly disclosing local data. However, current federated learning systems still face the dual threats of Byzantine attacks and user privacy leakage. The aim of this paper is to propose a privacy-preserving Byzantine robustness federated learning framework called PPBRFL, which achieves effective detection and defense against Byzantine attacks while protecting user data privacy. Specifically, PPBRFL performs data dimensionality reduction via randomized principal component analysis (RPCA) and combines the isolation forest algorithm with the amendable DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method for clustering to detect malicious clients comprehensively and accurately. In addition, a pseudo-gradient mechanism is introduced to prevent gradient leakage, thus ensuring model accuracy while realizing privacy protection. Experimental validation shows that PPBRFL can efficiently and accurately identify Byzantine attack nodes, effectively protect user data privacy, and exhibit excellent robustness.