Toward Provably Private Federated Learning
summary
Production systems claim formal user/device-level privacy by shipping DP-FTRL models for Gboard and running a large-scale secure aggregation SMPC stack; typical rounds use hundreds of thousands of clients and training takes thousands of rounds while Google runs billions of reports per day. A Trusted Computation Platform using AMD SEV-SNP TEEs, open source binaries and a Sigstore transparency log proves which code and DP steps run; empirical audits show lower memorization but SecAgg and TEEs remain vulnerable to Sybils and side channels.
AI-written summary · about
chapters (7)
- 0:00Federated learning intro
- 10:15Google FL evolution
- 23:51Differential privacy progress
- 32:02Verifiable privacy with TEE
- 44:22Private AI insights
- 50:18Open problems in FL privacy
- 57:50Q&A
related talks
14mToward Provably Private Insights into AI Use
32mBeyond Software: Hardware-Based Security
55mFederated, Synthetic, Personalized: Heterogeneity Here or There?
16mDP-BREM: Differentially-Private and Byzantine-Robust Federated Learning...
34mDecentralized Machine Learning
37mDemystifying Privacy Preserving Computing
38mDemystifying Privacy Preserving Computing
19mPrivate AI: Building Trust Through Verifiable Computation