ComputationIntegrityProof<T>

Models & Types · AiDotNet.FederatedLearning.Verification

Generates proofs of local training computation integrity (research-stage).

Inheritance: FederatedLearningComponentBase<T> → ComputationIntegrityProof<T>

Implements: IVerifiableComputation

For Beginners

This is the “holy grail” of verifiable FL: proving that a client actually ran the correct training algorithm (e.g., SGD for N epochs on its dataset) without revealing the data or the model weights.

How It Works

How it works (conceptually):

Current limitations: Full computation proofs for deep neural networks are extremely expensive. Current research (zkRNN 2026, ZKML Survey 2025) shows:

This implementation provides a simplified proof-of-concept using hash chains to verify training step ordering, not the full SNARK-based approach.

Reference:

Constructors

Constructor Summary
ComputationIntegrityProof(IZkProofSystem proofSystem, int expectedEpochs) Initializes a new instance of ComputationIntegrityProof.

Properties

Property Summary
string ProofSystemName { get; }

Methods

Method Summary
VerificationProof GenerateProof(byte[] value, VerificationConstraint constraint)
VerificationProof GenerateTrainingProof(TrainingStepLog log, int clientId) Generates a proof of training integrity from a training log.
bool Verify(VerificationProof proof, VerificationConstraint constraint)
void RecordTrainingStep(TrainingStepLog log, byte[] modelStateHash, double loss, int epoch) Records a training step by hashing the current state (model hash + loss + epoch).