InfluenceFunctionUnlearner<T>
Models & Types · AiDotNet.FederatedLearning.Unlearning
Influence function-based unlearning: mathematically estimates and subtracts a client’s contribution.
Inheritance: FederatedLearningComponentBase<T> → InfluenceFunctionUnlearner<T>
Implements: IFederatedUnlearner<T>
For Beginners
Influence functions answer the question: “How much would the model change if we removed one client’s data?” Instead of actually retraining, we mathematically estimate the answer using the Hessian (second derivative) of the loss. It’s like calculating how much a building would shift if one support column were removed, without actually removing it.
How It Works
How it works:
Trade-off: More accurate than gradient ascent for small removals (1-2 clients), but the Hessian approximation degrades for large removals or highly non-convex models.
Constructors
| Constructor | Summary |
|---|---|
InfluenceFunctionUnlearner(FederatedUnlearningOptions options) |
Initializes a new instance of InfluenceFunctionUnlearner whose curvature is the empirical Fisher (Gauss-Newton) matrix of the retained clients’ updates: H v = (1/M) sum_u u (u^T v). |
InfluenceFunctionUnlearner(FederatedUnlearningOptions options, Func<Tensor<T>, Tensor<T>, Tensor<T>> hessianVectorProduct) |
Initializes a new instance of InfluenceFunctionUnlearner with the exact Hessian of the retained loss, given as Hessian-vector products (for example double backpropagation through the gradient tape, or Pearlmutter’s R-operator). |
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) |