Privacy-Preserving Robust Federated Learning Based on Threshold Homomorphic Encryption

Abstract

Federated learning (FL) enables collaborative model training without sharing raw data, but remains vulnerable to Byzantine attacks and privacy leakage. Existing privacy-preserving robust FL schemes suffer from prohibitive computation and communication overheads, particularly on resourceconstrained clients. To address these challenges, we propose a lightweight dual-server robust FL framework based on (2,2)-threshold CKKS homomorphic encryption. Clients only perform local training, gradient clipping, and one-time encryption, while the dual servers collaboratively extract two-dimensional encrypted features and apply HDBSCAN clustering to isolate malicious clients. By shifting complex cryptographic operations to the dual-server side, the framework significantly reduces clientside overhead while maintaining high communication efficiency. Experiments show that the proposed framework achieves strong Byzantine robustness and high model accuracy while significantly outperforming existing schemes in computational and communication efficiency.