Robust Subgraph Learning by Monitoring Early Training Representations

Graph neural networks (GNNs) have attracted significant attention in security and artificial intelligence for their outstanding performance in graph learning and node classification tasks. However, GNN’s vulnerability to adversarial attacks, particularly through susceptible nodes, presents a major challenge for reliable decision-making. Addressing this challenge requires robust graph summarization techniques to mitigate the propagation of adversarial perturbations throughout the graph.In this paper, we propose SHERD (Subgraph Learning Hale through Early Training Representation Distances), a novel approach that enhances adversarial robustness while improving overall performance in graph inputs. SHERD leverages representations from a partially trained graph convolutional network (GCN) to detect susceptible nodes during adversarial attacks using standard distance metrics. By detecting and removing these "vulnerable" or "bad" nodes, SHERD constructs a robust subgraph that preserves and improves node classification performance. Through our experiments, we demonstrate the increased performance of SHERD in enhancing robustness by comparing the network’s performance on original and subgraph inputs against various baselines alongside existing adversarial attacks. Our experiments in multiple datasets, including citation datasets such as Cora, Citeseer, and Pubmed, as well as microanatomical tissue structures of cell graphs in the placenta, highlight that SHERD not only achieves substantial improvement in robust performance but also outperforms several baselines in terms of node classification accuracy and computational complexity. The implementation is available at https://github.com/sepideh-nsf/SHERD.