Reproduce “Toward provably private analytics and insights into GenAI use” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Albert Cheu, Artem Lagzdin, Brett McLarnon, Daniel Ramage, Katharine Daly, Marco Gruteser, Peter Kairouz, Rakshita Tandon and 3 more
Google deployed a system that provides provably private analytics and insights into Generative AI usage by integrating Trusted Execution Environments with lightweight LLM inference for unstructured data, ensuring verifiable confidentiality even from the server operator. This system processes billions of daily records and automates Differential Privacy parameter tuning.
Key insights
- Combining TEEs with a cryptographically-controlled Key Management Service and client-defined policies enables "provably private" analytics, verifiable even against a malicious server operator.
- It is feasible to run LLM inference engines inside TEEs to process and summarize sensitive unstructured data privately before differentially private aggregation.
- Automating Differential Privacy parameter tuning significantly improves the usability and adoption of privacy-preserving analytics without compromising guarantees.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.