Google Research Moves Federated Learning Into TEEs
Google Research integrates federated learning with server-side TEEs, deploying the system to Gboard for English and Japanese.
Stock photo for illustration only, not from the actual event
- Google Research shifts gradient computation from mobile devices to server TEEs
- Access policies are published to Sigstore's Rekor log for verifiable transparency
- Gboard currently utilizes the system for English and Japanese next-word prediction
Google Research has unveiled an advanced federated learning architecture that transitions gradient computation away from user smartphones and into attested server-side Trusted Execution Environments (TEEs). This architectural shift significantly hardens the privacy and security guarantees of machine learning models deployed at scale.
To ensure absolute accountability, access policies are officially published to Sigstore's Rekor log, while the corresponding binaries are engineered to be reproducibly buildable. This transparency allows external auditors and privacy advocates to independently verify central differential privacy mechanisms without compromising user data.
Stock photo for illustration only, not from the actual event
Real-world deployment is already underway, as Gboard has actively integrated this novel framework for next-word prediction features in both English and Japanese. Users benefit from enhanced predictive text capabilities backed by verifiable, externally auditable privacy protections.
Combining federated learning with TEEs addresses longstanding trust deficits in decentralized AI training. By leveraging cryptographic verification and reproducible builds, Google bridges the gap between server-side processing efficiency and stringent user privacy demands.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment