A Cross-Domain Adversarial-Resilient Graph Reduction Framework for Robust GNN-Based Node Classification

Abstract

Graph Neural Networks (GNNs) have demonstrated good results in graph structured data applications like node classification and representational learning. Although these models are successful, recent research demonstrates that these models are very susceptible to adversarial attacks. Even minor changes in the structure of the graph can cause a significant decrease in the prediction accuracy. Making GNN models more resilient to such attacks is thus also a significant research problem, particularly in systems where adversarial manipulations may influence information flow and potential attack propagation across graph structures.This paper proposes a robust graph reduction framework that enhances adversarial resilience by combining gradient-based node importance estimation with graph structure reduction. To determine the node significance, gradient sensitivity is used to identify the graph components that have a high impact on model predictions and information propagation during the GCN's forward pass. According to these importance scores, a reduced subgraph is created by keeping important nodes and eliminating less significant graph elements. The Graph Convolutional Network (GCN) is then trained on the reduced graph to evaluate whether emphasizing nodes the model has learned to rely on enhances adversarial resilience.The proposed framework is evaluated on the Cora citation network benchmark. Experimental results show that the reduced graph achieves higher clean classification accuracy than the full graph, while substantially reducing performance degradation under adversarial edge attacks. To assess generalization, the same gradient-based scoring approach is applied to three additional domains: a social network, a protein-protein interaction network, and a financial fraud detection network. Across all four domains, the reduced graphs achieve equal or improved accuracy relative to their respective baselines, with the retention budget empirically tuned per domain.