Disparate Vulnerability in Link Inference Attacks against Graph Neural Networks

Graph Neural Networks (GNNs) have been widely used in various graph-based applications. Recent studies shown that GNNs are vulnerable to link-level membership inference attacks (LMIA) which can infer whether a given link was included the training graph of GNN model. While most focus on privacy vulnerability links entire graph, none inspected risk specific subgroups (e.g., between LGBT users). In this paper, we present first study disparity subgroup (DSV) against LMIA. First, with extensive empirical evaluation, demonstrate existence non-negligible DSV under settings models and input graphs. Second, by both statistical causal analysis, identify difference three structural properties as one underlying reasons for DSV. Among properties, density has largest effect Third, inspired design new defense mechanism named FairDefense mitigate while providing protection At high level, at each iteration target model training, randomizes edges probability, aiming reduce gap different mitigation. Our results outperforms existing methods trade-off accuracy. More importantly, it offers better