source: Google Research: Toward provably private learning from federated data
level: technical
google research announced a new federated learning system that uses trusted execution environments to give externally verifiable privacy guarantees. the system shifts computation from devices to the server, improving training speed, accuracy, and device coverage. gboard has already adopted it for english and japanese next word prediction models. the design builds on four privacy principles: data minimization, anonymization, transparency, and verifiability. this replaces earlier systems that relied on trust in the server operator.
the new system encrypts training examples on devices and only decrypts them inside tee workloads that match pre-authorized access policies published to a public transparency log. a key management system running in tees enforces these policies. only metrics and differentially private model weights are visible to operators. gboard training time dropped from one to two months to substantially faster, with noise multipliers reduced by optimizing device participation schedules. limitations include current tee hardware side-channel risks and resource availability constraints.
the system uses an open-source orchestration language called federated language and supports reproducible builds from public code. external auditors can verify the exact code run on servers. this work is a step toward rigorous proof that server-side processing preserves individual privacy. future work includes integrating tees with accelerators for larger models and exploring synthetic data generation and llm inference workloads. the approach may one day include full proofs of correctness for differential privacy implementations.
why it matters: it lets organizations train models on private user data with externally auditable privacy guarantees, reducing the need to trust a central server operator.
source: Google Research: Toward provably private learning from federated data