Google Research has launched a new federated learning system that trains its Gboard keyboard models inside sealed server hardware. The announcement came on October 2, with a technical paper on arXiv.

Federated learning trains AI models using data from many phones without collecting raw data in one place. Google introduced the idea in 2017. The weakness was verification: users had to trust Google’s word on how servers handled the data.

In the new design, phones upload encrypted typing examples. Only sealed hardware, called Trusted Execution Environments, can decrypt them. The rules for which programs may touch the data are published in a public log called Rekor, so outside auditors can check them.

Google says the system cut training time for the English next-word model from about two months to three weeks. It is already used for Gboard’s English and Japanese prediction models.

The framework has been released as open source under the Apache 2.0 licence. Other developers can now build similar privacy-first training systems.