DP-BREM - Differentially-Private and Byzantine-Robust Federated Learning with Client Momentum

Learn about a novel federated learning protocol that simultaneously achieves differential privacy and Byzantine robustness through client momentum in this 16-minute conference presentation. Discover how DP-BREM addresses the vulnerability of existing federated learning systems to attacks that compromise both data privacy and model robustness by introducing a client momentum mechanism that averages updates over time to reduce honest client variance while exposing malicious perturbations from Byzantine clients. Explore the theoretical foundations of adding noise to aggregated momentum rather than gradients, and understand how the enhanced DP-BREM+ solution eliminates the need for a trusted server through secure aggregation techniques where differential privacy noise is jointly generated by clients. Examine comprehensive theoretical analysis and experimental results demonstrating superior privacy-utility tradeoffs and stronger Byzantine robustness compared to baseline methods across various differential privacy budgets and attack scenarios, making this essential viewing for researchers and practitioners working on secure and private machine learning systems.