Toward Provably Private Insights into AI Use
summary
A system called Provably Private Insights combines trusted execution environments with a “data expert” LLM and differential privacy to ensure raw user data never leaves enclaves and released aggregates carry a formal guarantee. Open-source code (Google Parfait) and a transparency log enable external attestation. Deployed in Pixel Recorder to classify transcripts: ran Gemma 3.4B inside TEEs, produced DP histograms with epsilon=1 from up to 100k inputs. Works end-to-end but CPU memory and throughput limit scaling to larger models (e.g., 20B) without accelerators.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:54Motivation for private AI
- 2:41Defining provable private insights
- 3:25Three pillars
- 6:57System architecture
- 10:26Pixel Recorder app
- 12:33Scalability and future work
related talks
01h 01mToward Provably Private Federated Learning
16mPrvTel: Lightweight Models for Private and Accurate Telemetry Data Retention
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
28mWhy Private AI Matters: Data Sovereignty as the Foundation for Trustworthy AI on the Open Web
19mPractical Considerations for Differential Privacy
01h 20mUnlocking societal benefits of AI with privacy protective technologies.
14mEstablishing Privacy Metrics for Genomic Data Analysis
50mThe ChatGPT Privacy Tango: Dancing with Data Security and Large Language Models