Google ships TEE-based federated learning, deployed on Gboard

Google released a new federated learning system built on trusted execution environments, already used to train Gboard's English and Japanese next-word prediction models. Encrypted training data processing and gradient computation move to the server side, with public access policies, remote attestation and reproducible builds for outside audit. Operators see only metrics and model weights that meet differential privacy requirements. Training that once took one to two months per model now runs in parallel across server machines, though TEE resource availability still caps it.

Written by AI from the coverage · updated 1 hour ago

Not enough continuous observations to draw a trend yet.