Google Research Moves Federated Learning Into TEEs: Gboard Now Trains With Externally Verifiable Differential Privacy
Google just moved federated learning gradient computation into attested TEEs. Gboard's next-word prediction now trains with externally verifiable differential privacy—binaries reproducible, access policies logged to Sigstore.
Why it matters
This shifts federated learning's privacy guarantees from 'trust us' to 'verify it yourself.' Practitioners deploying on-device ML can now audit central differential privacy via public logs and reproducible builds—a meaningful step toward transparent privacy in production systems.
The key facts
6 to knowGradient computation moved from phones to server-side TEEs (trusted execution environments)
Differential privacy applied centrally, not on-device
Access policies published to Sigstore's Rekor immutable log
Binaries reproducibly buildable for independent verification
Live deployment: Gboard English and Japanese next-word prediction
Externally auditable privacy—no manual trust required
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Google Research has unveiled a federated learning system that moves gradient computation from phones into attested server-side TEEs. Access policies are published to Sigstore's Rekor log and the binaries are reproducibly buildable, so central differential privacy can be checked externally. Gboard…