Google Secured Federated Learning With TEEs

A new system moves model training into secure server environments to provide externally verifiable privacy.

Updated on Oct. 4, 2026 in Artificial Intelligence

Bold flat-color editorial illustration in navy, cream, and red, featuring geometric shapes that symbolize secure hardware-level data encryption. — Google Research has deployed a new federated learning architecture that uses hardware-level Trusted Execution Environments to verify privacy protocols for model training. AI Illustration. Upload story photo >

Live Poll

Do you trust large technology companies to protect your personal data during artificial intelligence model training?

Google Research has introduced a federated learning architecture that leverages Trusted Execution Environments (TEEs) to enforce privacy protocols. The system is now being used to train Gboard next-word prediction models for English and Japanese users.

Why it matters

This development addresses a fundamental trust issue in federated learning where operators previously required oversight to ensure differential privacy noise was applied correctly. By using hardware-level security, the system allows for external verification of privacy guarantees.

The system utilizes RAFT for consensus in key management and Rekor logs for access policy transparency. It enables externally verifiable differential privacy across a cohort of 6,500 devices over 5,000 training rounds, a significant shift from the typical 1 to 2 month training period.

The players

Google

A multinational technology conglomerate that develops the Android ecosystem, machine learning frameworks, and Gboard.

Google Research

The internal division focused on advancing artificial intelligence and machine learning infrastructure.

The details

The architecture migrates client gradient computation to a server-side TEE—a secure hardware enclave that isolates data from the host operating system. Devices encrypt training examples locally and assign an access policy that governs TEE processing. Key management is coordinated via the RAFT consensus protocol—an algorithm used to ensure all nodes in a cluster agree on the same state—while access policies are validated using Rekor, a public transparency log.

2017: Google introduced the initial Federated Learning framework.
October 4, 2026: Google Research announced the TEE-based system.

The Tech Race

This move marks a departure from standard federated learning models that rely on implicit trust in the server operator. It aligns with broader industry efforts to harden AI infrastructure against unauthorized data access through hardware-backed enclaves.

Users of Gboard in English and Japanese are the first to experience this updated training workflow. The system is published under an Apache 2.0 open-source license, allowing developers to integrate these verifiable privacy protocols into their own federated learning applications.

The takeaway

The system represents a move toward verifiable, hardware-enforced privacy in decentralized machine learning pipelines. Watch for the adoption of these TEE architectures in broader industry standard benchmarks for privacy-preserving AI.