IGradientCommitment<T>
Interfaces · AiDotNet.FederatedLearning.Verification
Defines the contract for committing to gradient values before revealing them.
Known implementations: HashCommitmentScheme<T>, PedersenCommitment<T>
For Beginners
In FL, a malicious client could wait to see other clients’ updates and then craft its own update to manipulate the result (an “adaptive attack”). Commitment schemes prevent this: each client first sends a “commitment” (a cryptographic lock on its gradient), and only after all commitments are received does anyone reveal the actual values.
How It Works
Properties:
Homomorphic commitments (like Pedersen) have an additional property: the server can verify that the sum of committed values matches the committed sum, without seeing individual values.
Methods
| Method | Summary |
|---|---|
GradientCommitmentData<T> Commit(Tensor<T> gradient) |
Creates a commitment to a gradient tensor. |
Tensor<T> Open(GradientCommitmentData<T> commitment) |
Opens a commitment, revealing the original gradient value. |
bool Verify(GradientCommitmentData<T> commitment) |
Verifies that an opened commitment matches the original commitment. |
bool VerifyAggregation(IReadOnlyList<GradientCommitmentData<T>> individualCommitments, GradientCommitmentData<T> aggregateCommitment) |
Verifies that the sum of individual commitments matches a claimed aggregate commitment. |