Title: Bayesian Membership Privacy for Graph Neural Networks

Abstract:

Current privacy evaluations for Graph Neural Networks (GNNs) frequently rely on assumptions derived from non-graph contexts, thereby neglecting structural dependencies and the stochastic nature of training-graph sampling. Specifically, because priors vary by node, relying solely on type-I and type-II errors fails to adequately define the optimal membership inference test. To remedy this limitation, we propose Bayesian Membership Privacy (BMP), a formulation for node-level privacy that accounts for sampling. BMP integrates node-specific priors and considers graph sampling probabilities as integral to the adversary’s knowledge base. By framing membership inference as a Bayesian hypothesis test, BMP quantifies privacy through posterior membership probability. We analyze the theoretical attributes of BMP in comparison to established definitions in existing literature. Additionally, we introduce a practical, sampling-aware auditing framework designed to estimate BMP parameters, thereby serving as a metric for node-level privacy leakage within GNNs. Our experiments on standard benchmark graph datasets demonstrate that BMP provides granular privacy insights that remain obscured when relying exclusively on global attack accuracy.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC