Google Research has launched a next-generation federated learning system built on trusted execution environments, bringing externally verifiable central differential privacy guarantees to machine learning for the first time. The updated architecture shifts client gradient computation directly into attested server environments, closing a long-standing trust gap where users and developers had to rely solely on central operators to apply differential privacy correctly and handle sensitive data securely.
The system coordinates four core operational components to ensure verifiable data handling. Devices locally encrypt training examples and pre-authorize strict access policies that are publicly recorded in the Rekor transparency log. A key management system built from trusted execution environments running the RAFT consensus protocol releases decryption keys exclusively to workloads matching these verified policies. A root execution environment then manages a Python training loop using frameworks derived from TensorFlow Federated, releasing only differentially private model weights while ensuring encrypted data can only be accessed for a strictly limited duration.
Google has already deployed this new framework in production to power English and Japanese next-word prediction models on Gboard, achieving stronger privacy protections alongside improved model accuracy. By moving the training bottleneck from mobile devices to parallelized server infrastructure, the system eliminates traditional delays caused by diurnal swings in device availability and substantially accelerates compute times. This development matters because it establishes a transparent, audibly verifiable standard for privacy-preserving machine learning that eliminates the need for blind trust in centralized sy