DP-BREM: Differentially-Private and Byzantine-Robust Federated Learning...
summary
Cross-silo federated learning uses client momentum that averages each client’s updates over rounds so tiny malicious perturbations accumulate and become detectable, then adds calibrated DP noise to the aggregated momentum rather than to per-update gradients. The protocol DP-BREM claims record-level differential privacy with a tighter sensitivity analysis than DP-SGD, and DP-BREM+ replaces a trusted server by having clients jointly generate DP noise via secure aggregation/verifiable sharing. Experiments on standard benchmarks show improved accuracy under increasing percentages of Byzantine clients and better privacy–utility tradeoffs than baselines.
AI-written summary · about
chapters (14)
- 0:00Federated learning threats
- 3:02Differential privacy basics
- 5:32System model and threats
- 6:16DP and robustness challenges
- 7:21DP-BREM approach
- 8:02History momentum method
- 9:07Centralized DP noise strategy
- 9:49DP-BREM workflow
- 11:14DP-BREM Plus secure
- 11:57DP-BREM privacy
- 12:39Robustness and convergence
- 13:24Robustness experiments
- 14:07Privacy-utility trade-off
- 15:11Summary and future
related talks
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms
01h 01mToward Provably Private Federated Learning
33mPrivacy of Decentralized Machine Learning
58mLarge-Scale Private Learning, Part III
48mSynthetic Data as an Enabler for Learning from Decentralized Private Data
33mHeterogeneity and Privacy in Modern Learning
31mScaling privacy embedded distributed ML systems
37mUser-level and federated local differential privacy