Safety, alignment & governance
Toward provably private learning from federated data
The system uses Trusted Execution Environments (TEEs) to make server-side federated learning workloads more verifiable, while limiting which code can access encrypted training data.
Why it matters: The design lets auditors inspect published workloads and limits operators’ visibility to metrics and differentially private model weights. It also shifts training bottlenecks from device availability and compute to server-side TEE resources.