From Decoupling to Reconstruction: A Robust Graph Neural Network Against Topology Attacks

Graph neural networks (GNNs) are criticized for poor robustness; their performance will decline seriously even with unnoticeable edge perturbations in node classification tasks. To against topology attacks, recent works punish edges based on data and attack characteristics (such as node similarity or SVD), or utilize the graph data distribution to absorb attacks (such as BayesianGCN or RGCN). We propose a robust GNN model based on topology reconstruction, which does not explicitly assume data or attack distributions. Our proposed method decouples graph data into attribute subgraph and topology subgraph and makes classifier cooperate with topology generator to label nodes. The classifier starts training from the attribute subgraph and labels some high-confidence nodes to join the trainset. Meanwhile, the generator reconstructs topology based on classification loss, topology subgraph, and link prediction, providing the classifier with new information. Extensive node classification experiments show that under the same attack conditions, our model has higher classification accuracy and stronger robustness compared with other latest robust GNN models.