Google Unveils Groundbreaking Federated Learning System with Verifiable Privacy via TEEs

Revolutionizing Privacy in AI Training

Google Research recently announced a significant advancement in Federated Learning (FL), unveiling a novel system that harnesses Trusted Execution Environments (TEEs). This development is notable for its claim to be the first to offer externally verifiable central differential privacy (DP) assurances within an FL framework, addressing long-standing challenges in data trust and transparency for AI models.

Addressing Previous Federated Learning Gaps

Since its inception in 2017, Federated Learning has powered numerous smart features, including next-word prediction in Gboard, reply suggestions in Google Messages, and intelligent text selection on Android. However, earlier iterations faced a fundamental trust deficit: external parties lacked mechanisms to confirm that user data was neither logged nor inadvertently exposed during aggregation processes. While cryptographic secure aggregation provided some protection, it proved incompatible with advanced central differential privacy algorithms. This also meant relying on the platform operator to correctly apply privacy noise.

The newly designed system fundamentally shifts how privacy is managed. Instead of client devices performing gradient computations, this work moves to server-side TEEs. Critically, the logic executed within these server environments can be attested, removing the need for users to implicitly trust the operator's practices.

How the New System Functions

Building upon previous confidential federated analytics efforts, the enhanced system coordinates four essential components:

• Encrypted Data Uploads: Devices locally encrypt training data and pre-authorize specific access policies, which dictate which TEE computations are permitted to process the information. These policies are transparently published in a public log.
• Policy-Driven Key Management: A Key Management System, comprised of TEEs operating under the RAFT consensus protocol, selectively issues decryption keys exclusively to server-side workloads whose attested code precisely matches the pre-defined access policy.
• Secure Workload Execution: A root TEE orchestrates a Python training loop, distributing parallel subtasks to worker TEEs. This process, facilitated by the open-source Federated Language (derived from TensorFlow Federated), ensures that only differentially private model weights are ultimately released to data analysts.
• Robust Fault Tolerance: Following each training round, the system saves a KMS-encrypted recovery state. This safeguard enables seamless recovery from root or worker failures without compromising privacy-sensitive information.

Establishing Verifiable Privacy

The system's privacy guarantees are externally verifiable through several mechanisms. Access policies are recorded in Rekor, Sigstore's public transparency log, allowing external auditors to meticulously track and review every server workload that could potentially process device data. Furthermore, the Key Management System and data processing binaries are reproducibly buildable from their open-source codebases, offering an additional layer of assurance. While TEEs accommodate sideloading serialized logic to protect proprietary model architectures, all privacy-critical logic remains hardcoded within the attested program. Operators are restricted to viewing only aggregate metrics and differentially private model weights, with encrypted data decrypted only for a strictly limited duration post-upload.

Transforming Gboard's Performance

This innovative system has already been deployed for Gboard's English and Japanese next-word prediction models, yielding both stronger privacy assurances and enhanced accuracy. This improvement stems from two key design decisions:

• Centralized Data Collection: All data uploads are now collected before server-side training commences. This eliminates slowdowns previously caused by diurnal fluctuations in device availability, enabling the program to compute optimal participation schedules and finely tune differential privacy parameters for improved utility.
• Server-Side Scalability: The primary computational bottleneck has been shifted to the server. Unlike prior FL models that often required one to two months for training, the new system allows training to be parallelized across numerous machines, limited only by available TEE resources. This translates to substantially faster computation times, although specific speedup figures have not been disclosed.