Google moves Gboard model training to servers with auditable privacy controls

Google announced a federated learning system already used for Gboard's English and Japanese next-word prediction models that shifts training computation from devices to servers. Google says the system improves accuracy and training speed while allowing external auditors to verify privacy controls. Encrypted device data can be processed only inside protected computing environments running programs specified in published access policies, and only for a limited time after upload. Operators see metrics and model weights with privacy-preserving noise. Public logs identify permitted workloads, and core binaries can be rebuilt from open source code. Google says training is now limited by protected server resource availability; current hardware limitations and side-channel observations remain areas for further protection.