Secure-FL: PyTorch clients around a Rust round server Diagram

No framework in the middle

Clients train in PyTorch and talk gRPC; there is no in-process FL framework
The round server samples clients, closes a round when every assigned client submitted, and publishes the model with its proof
Proofs, aggregation and the stability heuristic all run in the same Rust core

Verification gate

An update is dropped unless its state hash, delta hash, learnable norm and Groth16 proof all pass
The norm bound comes from the model, not from the client's own claim
The server self-checks every aggregation proof before releasing the model

Identity

Writes need a bearer token; mTLS makes the certificate CN the client identity
A valid certificate cannot submit under another client_id
REST stays as a debug front-end and does not bind certificate identity