GraphNeighborhoodPrivacy<T>
Models & Types · AiDotNet.FederatedLearning.Graph
Applies local differential privacy (LDP) to neighborhood queries to prevent topology leakage.
Inheritance: FederatedLearningComponentBase<T> → GraphNeighborhoodPrivacy<T>
For Beginners
GNN embeddings encode information about a node’s neighborhood structure. If shared naively, an adversary could reconstruct the local graph topology. This class adds calibrated noise to neighborhood aggregations before they leave a client, ensuring that individual edges cannot be inferred from the shared embeddings.
How It Works
Mechanism:
Privacy guarantee: (epsilon, delta)-differential privacy for individual edges. Lower epsilon = stronger privacy but more noise.
Constructors
| Constructor | Summary |
|---|---|
GraphNeighborhoodPrivacy(double epsilon, double delta, double sensitivity) |
Initializes a new instance of GraphNeighborhoodPrivacy. |
Methods
| Method | Summary |
|---|---|
Tensor<T> ClipAndNoiseNeighborhood(Tensor<T> aggregatedFeatures, double clipNorm) |
Clips and adds noise to an aggregated neighborhood feature vector. |
Tensor<T> PrivatizeEmbeddings(Tensor<T> embeddings, int embeddingDim) |
Adds calibrated Gaussian noise to node embeddings to protect neighborhood structure. |
bool RandomizedResponseEdge(bool edgeExists) |
Applies randomized response to edge existence queries. |
double[] PerturbDegrees(int[] degrees) |
Perturbs node degree information with Laplace noise. |