Google Implemented New Federated Learning System

The company has launched an AI training model that uses trusted execution environments for improved privacy.

Updated on Oct. 4, 2026 in Artificial Intelligence

Isometric editorial illustration of stacked matte rectangular blocks, representing secure data processing architecture in an institutional blue and oxblood palette. — Google has deployed a new federated learning system for Gboard, utilizing server-side Trusted Execution Environments to provide verifiable privacy protections for next-word prediction. 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 deployed a federated learning system that utilizes Trusted Execution Environments to verify differential privacy. This technology is currently powering next-word prediction models for Gboard users in English and Japanese.

Why it matters

The system eliminates the need for users to trust the operator to apply privacy protections. By using server-side execution environments, it provides externally verifiable privacy guarantees that were previously unavailable.

The system processes training using Python loops and Federated Language, while relying on the RAFT consensus protocol for key management. It allows for training in 5,000 rounds compared to previous timelines of 1 to 2 months.

The players

Google Research

This is the advanced technology division of Google that focuses on long-term computer science projects and artificial intelligence development.

Gboard

This is the virtual keyboard application developed by Google for Android and iOS mobile devices.

The details

Devices encrypt training data locally and authorize an access policy for server-side Trusted Execution Environments, which handle gradient computation. Google has released the framework under an Apache 2.0 open-source license to support transparency via public logs.

Google introduced its initial Federated Learning framework in 2017.
The new TEE-based system was officially announced on October 4, 2026.

The Tech Race

This development represents a shift from trust-based privacy models to mathematically verifiable security in distributed AI training. It builds upon Google's 2017 Federated Learning framework to address vulnerabilities inherent in earlier operator-dependent systems.

Users will benefit from enhanced privacy protections when their devices contribute to next-word prediction improvements. The implementation ensures that personal training data remains encrypted while still allowing the keyboard to learn from local language patterns.

The takeaway

This technology demonstrates that machine learning can evolve without sacrificing individual user data privacy. Developers and researchers can utilize the open-source code to build similar privacy-first training environments for other applications.