Google's New Federated Learning System Trains Models in TEEs

Google has announced a new federated learning system that trains models inside trusted execution environments (TEEs) on its servers. This system was revealed in a blog post and paper published by Google Research on October 2, 2026. According to the paper, the system allows for faster training times and improved privacy compared to traditional federated learning methods.

The new system uses TEEs to train models on user data without exposing sensitive information. In one experiment, the system trained a Gboard English model in just three weeks, whereas the previous production model took two months to train. The system also achieved a 3x smaller privacy budget compared to the baseline system.

Another experiment demonstrated improved device coverage using the Japanese model. The system was able to collect contributions from 17.8 million devices over six days, which is about 2.1 times as many devices as the baseline system collected in 38 days.