Toward provably private learning from federated data

Google says its new federated learning system lets outsiders verify the privacy protections around server-side training.

The system uses trusted execution environments, public access policies, a KMS, and reproducible open-source binaries to restrict how uploaded training data can be decrypted and processed. Google says only metrics and differentially private model weights are visible to workload operators. Gboard has already used it for English and Japanese next-word prediction models, with faster training and improved accuracy. Google Research's note