Efficient Robust Private Aggregation for DP
summary
Two-server Prio model for secure aggregation, with concrete constructions: PINE: sqrt(D) additional communication, about 3% overhead at D=1,000,000, gives robust L2 norm verification via a dot-product test with z in {-1,0,1} and a small-interval finite-field check, and PREAMBLE: randomized block sampling plus DPFs yields roughly 100× lower client bandwidth with sublinear communication and server O(D) work while preserving near-standard DP noise. Uses random rotations, block truncation to avoid large truncation error, and privacy amplification by (balls-and-bins) sampling.
AI-written summary · about
chapters (9)
- 0:00Private aggregation problem
- 2:33Prio model and secure aggregation
- 7:38Robustness challenges in Prio
- 9:48Handling real numbers in finite fields
- 12:42PINE protocol for Euclidean norm
- 18:25Sublinear communication with Preamble
- 24:51Block-based sampling for efficiency
- 30:43Privacy error and performance results
- 32:55Summary and future directions
related talks
01h 07mIndustry Presentations
44mFederated Heavy Hitters Discovery with Differential Privacy
14mDistributed Private Aggregation in Graph Neural Networks
01h 10mDifferential Privacy: Fundamentals to Forefront II
37mUser-level and federated local differential privacy
09mFrom Robustness to Privacy and Back
16mDP-BREM: Differentially-Private and Byzantine-Robust Federated Learning...
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms