Beyond Homophily: Towards Generalized Graph Reconstruction Attack and Defense
Researchers have developed new methods for attacking and defending graph neural networks (GNNs) against information leakage. The study characterizes how graph properties like homophily and heterophily influence the recoverability of training data. Building on a Markov chain approximation, they propose an attack that reconstructs graph adjacency by aligning representations across GNN layers and a defense that suppresses this sensitive information while maintaining classification accuracy. AI
IMPACT Introduces new techniques for privacy preservation in GNNs, potentially impacting how sensitive graph data is handled.