Bayesian Membership Privacy for Graph Neural Networks
Researchers have introduced Bayesian Membership Privacy (BMP), a new framework for assessing privacy in Graph Neural Networks (GNNs). BMP accounts for structural correlations and stochastic training-graph sampling, which are often overlooked in existing privacy analyses. By framing membership inference as a Bayesian hypothesis test, BMP quantifies privacy based on posterior membership probability and offers a more granular insight into privacy leakage than global attack accuracy. AI
IMPACT Introduces a novel privacy auditing mechanism for GNNs, potentially improving data protection in graph-based AI applications.