Federated Retraining

Every ~10,000 rounds, validators collaboratively update models.

🔄 Steps:

1. Compute local gradients from recent (features, truth) samples

2. Share gradients via MPC

3. Aggregate into new model weights

4. Generate zk-SNARK that loss is minimized

5. Submit:

function updateModel(bytes32 newHash, bytes calldata proof) external;

1. Nodes pull updated weights automatically

🤝 Keeps consensus models adaptive while preserving privacy and integrity.