GradientAscentUnlearner<T>
Models & Types · AiDotNet.FederatedLearning.Unlearning
Approximate unlearning via gradient ascent: reverses learning by ascending the loss on target data.
Inheritance: FederatedLearningComponentBase<T> → GradientAscentUnlearner<T>
Implements: IFederatedUnlearner<T>
For Beginners
Normal training minimizes the loss (gradient descent — going downhill). Gradient ascent does the opposite: it maximizes the loss on the target client’s data, effectively making the model “forget” what it learned from that client. Think of it as deliberately making the model bad at the target client’s data while keeping it good at everyone else’s.
How It Works
How it works:
Speed: Much faster than exact retraining (minutes vs. hours). Provides approximate guarantees — the target client’s influence is reduced but not provably zero.
Constructors
| Constructor | Summary |
|---|---|
GradientAscentUnlearner(FederatedUnlearningOptions options) |
Initializes a new instance of GradientAscentUnlearner. |
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) |