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

Area of Science:

• Machine Learning
• Cybersecurity
• Distributed Systems

Background:

• Federated Learning (FL) enables collaborative model training but faces vulnerabilities to privacy and robustness attacks.
• Existing defenses often prioritize either privacy or robustness, not both.
• Cross-silo FL settings require robust solutions against sophisticated threats.

Purpose of the Study:

• To develop Federated Learning protocols that simultaneously achieve differential privacy (DP) and Byzantine robustness.
• To address the limitations of existing FL defenses that do not offer combined privacy and robustness guarantees.
• To propose methods resilient to attacks in cross-silo FL environments.

Main Methods:

• Introduced client momentum to average updates over time, enhancing robustness against Byzantine attacks.
• Developed DP-BREM, applying DP by adding noise to aggregated momentum and accounting for its privacy cost.
• Proposed DP-BREM+ using secure aggregation for decentralized DP noise generation, eliminating the need for a trusted server.

Main Results:

• DP-BREM and DP-BREM+ demonstrate superior privacy-utility tradeoffs compared to baseline methods.
• The proposed protocols exhibit enhanced Byzantine robustness under various attack scenarios.
• Theoretical analysis and experiments validate the effectiveness of the developed methods.

Conclusions:

• The developed DP-BREM and DP-BREM+ protocols offer a robust solution for secure and private Federated Learning.
• Client momentum and secure aggregation are effective techniques for achieving simultaneous DP and Byzantine robustness.
• These findings advance the field of secure and reliable distributed machine learning.