Google unveils next-gen federated learning with stronger privacy

Google has launched a new version of its federated learning system—where machine learning models train across users' devices, instead of on centralized data—that now uses Trusted Execution Environments (TEEs) for much stronger privacy and verifiability. TEEs let third parties verify computation and keep data confidential. Google says Gboard is already using this system, which also speeds up training time compared to previous versions.

• Uses Trusted Execution Environments (TEEs) for privacy
• Clients encrypt and control access to their device data
• Policy access must be published to a transparency log
• Gboard's adoption showed faster computation speeds
• System provides end-to-end verifiable privacy guarantees