Robust and Secure Federated Learning With Verifiable Differential Privacy

Federated Learning (FL) has emerged as a popular paradigm for training machine learning models on data from multiple sources without the need for data to leave their respective locations. Although FL is widely used, recent research indicates its vulnerability to data privacy breaches and Byzantine attacks. Addressing both threats simultaneously remains challenging, as previous defense mechanisms struggle to balance Byzantine resilience, data privacy, and training accuracy. Although combining Differential Privacy (DP) with Byzantine-resilience in FL offers potential solutions, it presents challenges due to theoretical incompatibilities that can impact the performance of large-scale models. Addressing this underexplored challenge, we propose a novel approach that enhances privacy and Byzantine-resilient in FL while maintaining training accuracy. Our method integrates DP with Byzantine-resilience techniques, overcoming the limitations of existing methods that often sacrifice accuracy or even Byzantine resistance. Our experiments carefully analyze the impact of DP noise on training accuracy and Byzantine-resilience, guiding appropriate parameter selection. The experimental results demonstrate that our approach achieves superior global model accuracy compared to previous competitive methods, even with strong privacy constraints.