This is the repository for the paper: Membership Inference Attack on Graph Neural Network https://arxiv.org/abs/2101.06570.
TSTF: Train on a subgraph and test on the full graph. The testing subgraph includes the full structure of the entire graph including the structure information of the training nodes.
TSTS: Train on a subgraph, test on another subgraph. Different subgraphs are used for training and testing the model. Only the structural infomation of selected graph is used during testing.
The model_type, data_type and defense_type can be changed to view different performance across different models and dataset.
Fig. 7. Relaxing the knowledge of the hyperparameter assumption. We varied the number of hidden neurons. The original
shadow model was trained with 256 hidden neurons.
Fig. 8. Relaxing the knowledge of target model. O = Original performance when target and shadow model have the same architecture. S = using SGC as shadow model, G = using GCN as shadow model. We observe a similar trend on the precision and recall.