RobustPPFL: A Secure and Robust Privacy-Preserving Federated Learning Framework Against Poisoning Attacks

Abstract

Federated learning (FL) facilitates decentralized machine learning by enabling participating entities to co-operatively train shared models while retaining local data ownership. To address privacy concerns inherent in distributed training, privacy-preserving mechanisms have been incorporated into FL frameworks to enhance confidentiality and protect sensitive data. However, the implementation of these privacy enhancing techniques introduces vulnerabilities to poisoning attacks, wherein malicious actors manipulate training processes to degrade model integrity or performance. Current defense strategies rely on statistical methods to mitigate such attacks, but they lack sufficient robustness, demonstrating limited effectiveness against diverse attack types or scenarios where malicious clients exceed a small minority. To overcome these limitations, we propose RobustPPFL, a privacy-preserving federated learning framework designed to withstand multiple poisoning attack types with high resilience, even when malicious participants dominate the client population. Our approach integrates three core innovations. First, we add performance verification of client-encrypted models during the training process to detect malicious clients in-time. Second, a secure model inference protocol is proposed to enable privacy-preserving training. Last but not the least, we design a grouped verification mechanism enhanced by hierarchical aggregation rules to optimize efficiency and minimize interference in malicious client detection. We evaluate RobustPPFL through extensive experiments across diverse datasets and attack scenarios. The experimental results show that our proposed framework achieves privacy preservation and it achieves high robustness against poisoning attacks.