A Verifiable, Privacy-Preserving, and Poisoning Attack-Resilient Federated Learning Framework
Washington Enyinna Mbonu, Carsten R. Maple, Gregory Epiphaniou, Christo Panchev · Big Data and Cognitive Computing · 2025
Federated learning is the on-device, collaborative training of a global model that can be utilized to support the privacy preservation of participants’ local data. In federated learning, there are challenges to model training regarding privacy preservation, security, resilience, and integrity. For example, a malicious server can indirectly obtain sensitive information through shared gradients. On the other hand, the correctness of the global model can be corrupted through poisoning attacks from malicious clients using carefully manipulated updates. Many related works on secure aggregation and poisoning attack detection have been proposed and applied in various scenarios to address these two issues. Nevertheless, existing works are based on the trust confidence that the server will return correctly aggregated results to the participants. However, a malicious server may return false aggregated results to participants. It is still an open problem to simultaneously preserve users’ privacy and defend against poisoning attacks while enabling participants to verify the correctness of aggregated results from the server. In this paper, we propose a privacy-preserving and poisoning attack-resilient federated learning framework that supports the verification of aggregated results from the server. Specifically, we design a zero-trust dual-server architectural framework instead of a traditional single-server scheme based on trust. We exploit additive secret sharing to eliminate the single point of exposure of the training data and implement a weight selection and filtering strategy to enhance robustness to poisoning attacks while supporting the verification of aggregated results from the servers. Theoretical analysis and extensive experiments conducted on real-world data demonstrate the practicability of our proposed framework.