Toward provably private analytics and insights into GenAI use

TL;DR

This paper addresses how to derive provably private analytics from unstructured GenAI usage while preventing access to raw data by the operator. It proposes a TEEs-based federated analytics architecture that uses a KMS inside TEEs, HPKE for secure data handling, and external verifiability via public transparency logs, delivering ($\varepsilon$, $\delta$)-DP guarantees on released results. A key contribution is the integration of on-device structured summarization using Gemma within TEEs, followed by server-side DP aggregation and autotuning of privacy parameters to preserve utility and security. The system scales to large data fleets, supports multiple workloads, and has production deployments, offering practical insights into GenAI usage with strong privacy guarantees and verifiability; future work includes accelerators, flexible DP units, and $k$-anonymity.

Abstract

Large-scale systems that compute analytics over a fleet of devices must achieve high privacy and security standards while also meeting data quality, usability, and resource efficiency expectations. We present a next-generation federated analytics system that uses Trusted Execution Environments (TEEs) based on technologies like AMD SEV-SNP and Intel TDX to provide verifiable privacy guarantees for all server-side processing. In our system, devices encrypt and upload data, tagging it with a limited set of allowable server-side processing steps. An open source, TEE-hosted key management service guarantees that the data is accessible only to those steps, which are themselves protected by TEE confidentiality and integrity assurance guarantees. The system is designed for flexible workloads, including processing unstructured data with LLMs (for structured summarization) before aggregation into differentially private insights (with automatic parameter tuning). The transparency properties of our system allow any external party to verify that all raw and derived data is processed in TEEs, protecting it from inspection by the system operator, and that differential privacy is applied to all released results. This system has been successfully deployed in production, providing helpful insights into real-world GenAI experiences.

Figures (2)

• Figure 1: The data and verification flow in our TEE-based private analytics system.
• Figure 2: An example differentially private distribution of transcripts in the Recorder application across various topics, as categorized by the Gemma model. Inner rectangle size is proportional to relative topic frequency.