Google Research announced a new federated learning system using Trusted Execution Environments. It provides externally verifiable privacy guarantees for data processing. The system shifts computation to the server to improve training speed and accuracy. Gboard has adopted it for next-word prediction models in English and Japanese. Training now completes in weeks instead of months. Device data is encrypted and only decrypted within TEEs matching published access policies. A public transparency log lets external auditors verify allowed workloads. The system supports fault-tolerant recovery without leaking sensitive data. It enables larger model training by removing on-device compute limits. Future work targets full proofs of correctness for DP algorithms and system components.
What changed
Federated learning relied on device-side computation and cryptographic aggregation without verifiable server-side privacy guarantees.
What it unlocks
Running federated training workloads with externally verifiable privacy guarantees on server-side TEEs.
- 5000 training rounds
- 6500 devices per cohort
- 1-2 months to 1-2 weeks training time
What you need to act on it
- TEE hardware availability
- access policy publication to a public transparency log
Sources