Google updated its federated learning system to improve privacy

Google released a technical whitepaper outlining a new federated learning system that moves training computations to the server while keeping user data private. This method relies on Trusted Execution Environments (TEEs) to ensure that the code processing data is both secure and auditable.

Federated learning allows machine learning models to learn from decentralized data without collecting the raw information itself. In this new design, devices encrypt their data before uploading it, and servers process this data only within secure containers. These containers verify exactly what logic is running, so outside parties can audit the process.

This system already powers Gboard features like next-word prediction in English and Japanese. Moving the heavy lifting to the server side helps solve speed issues caused by varying device availability.

How it works

• Devices encrypt data and pre-authorize specific server computations.
• A public log tracks these authorized workloads for transparency.
• TEEs ensure the server cannot see the raw data while it performs training.
• The system automatically recovers from potential failures without compromising privacy.

By using this setup, developers can train larger models because they are no longer limited by the processing power of individual phones. Have you ever wondered if your phone's predictive text is actually learning from your habits without keeping a copy of your messages?

This approach still faces limitations. It relies on the current capabilities of TEE hardware and requires ongoing research into mitigating side-channel attacks. The system currently does not offer full proofs of correctness for all software components, though the team expects this to be possible in the future.