DiffusiveNoiseUnlearner<T>
Models & Types · AiDotNet.FederatedLearning.Unlearning
Unlearning via structured diffusive noise injection targeting memorized samples.
Inheritance: FederatedLearningComponentBase<T> → DiffusiveNoiseUnlearner<T>
Implements: IFederatedUnlearner<T>
For Beginners
This approach (from 2025 research) works like adding carefully designed static to a radio signal to drown out a specific station. Instead of reversing the learning directly, we inject noise that is specifically structured to disrupt the model’s memorization of the target client’s data patterns. Then we “heal” the model by fine-tuning on remaining clients.
How It Works
How it works:
Advantages: More robust than gradient ascent (which can overshoot), and doesn’t require Hessian computation (which is expensive for large models). The structured noise targets memorization specifically rather than general model quality.
Constructors
| Constructor | Summary |
|---|---|
DiffusiveNoiseUnlearner(FederatedUnlearningOptions options) |
Initializes a new instance of DiffusiveNoiseUnlearner. |
Properties
| Property | Summary |
|---|---|
string MethodName { get; } |
Methods
| Method | Summary |
|---|---|
ValueTuple<Tensor<T>, UnlearningCertificate> Unlearn(int targetClientId, Tensor<T> globalModel, Dictionary<int, List<Tensor<T>>> clientHistories) |