Google Research Moves Federated Learning Into TEEs: Gboard Now Trains With Externally Verifiable Differential Privacy

Google Research has integrated its federated learning system with server-side Trusted Execution Environments to process Gboard training data. By publishing access policies to the Sigstore Rekor log, the company allows for external verification of differential privacy protocols. This shift moves gradient computation away from individual devices while ensuring that the central aggregation process remains transparent and auditable.