Private analytics via zero-trust aggregation

Details

Google Research has published a private analytics system designed to monitor on-device AI models — like those running in Android's SafetyCore — without compromising user privacy. The challenge: understanding whether a deployed model is drifting, biased, or failing requires aggregate data, but collecting per-user data creates privacy risk at scale.\n\nThe solution layers two mechanisms. First, a new cryptographic aggregation protocol that mathematically guarantees only anonymized, population-level insights reach Google — individual device data cannot be reconstructed. Second, Trusted Execution Environments (TEEs) provide hardware-backed attestation that the aggregation logic is running exactly as published, verifiable by external parties. This extends prior federated analytics work used in Pixel Recorder and Gboard.\n\nThe zero-trust design principle means no single entity can observe individual user data — including Google itself. The combination of cryptographic proof and hardware attestation sets a materially higher bar than typical federated learning implementations, and is relevant to any enterprise deploying on-device AI at population scale under privacy constraints.