Privacy Preserving Aggregation of Ad Conversions using MPC
summary
Prototype implements a three-party honest-majority MPC with malicious security (abort), using Boolean secret shares, vectorized bitwise operations, a shuffle, and distributed ZKPs; it converts to arithmetic only to evaluate the DY PRF for joins. The system processes 2 billion impressions and conversions across 150 shards (3 parties × 50 shards) in 1–2 hours, reveals match rates and per-bucket counts (device-side differential privacy for noise/budget), is open-source Rust, and costs about $0.30 per million impressions.
AI-written summary · about
chapters (10)
- 0:00Privacy preserving ad conversions
- 2:12Standard MPC limitations
- 4:43Match keys for secure attribution
- 9:59Malicious security design
- 15:16Joining reports with secret shares
- 17:43Scaling with MPC rings and sharding
- 21:39Engineering optimizations
- 28:05Protocol implementation challenges
- 30:11Results and cost efficiency
- 33:47Q&A
related talks
21mFrom Usability to Secure Computing and Back Again
49mCompiling to preserve our privacy
01h 31mIntroduction to Secure Multi-Party Computation
30mDistributed Aggregation Protocol: A Standard for Privacy-Preserving Aggregation in MPC
41mMPC-in-the-Head based Zero Knowledge Proofs
25mMCMix: Anonymous Messaging via Secure Multiparty Computation
17mSecure Vickrey Auctions for Online Advertising
23mRecent progress in MPC-in-the-Head protocols