Enhancing robust node classification via information competition: An improved adversarial resilience method for graph attacks

Abstract

Graph neural networks (GNNs) demonstrate their effectiveness in facilitating node classification and a range of graph-based tasks. However, recent studies have revealed that GNNs can be vulnerable to various adversarial attacks. Despite various defense strategies, ranging from attack-agnostic defenses to attack-oriented defenses that have been proposed to mitigate the impact of adversarial attacks on graph data, effectively learning attack-agnostic graph representation remains an open challenge. This paper introduces a novel information Competition-based framework for Graph Neural Networks (i.e., iC-GNN, e.g., iC-GCN, iC-GAT, etc.) to enhance the robustness of GNNs against various adversarial attacks in node classifications. Through the use of graph reconstruction and low-rank approximation, our approach learns diversified graph representations to collaboratively perform node classifications. Meanwhile, mutual information constraints are utilized on different graph representations to ensure diversity and competition in graph features. The experimental results indicate that within the proposed framework, iC-GCN outperforms other graph defense frameworks in countering a wide range of targeted and non-targeted adversarial attacks in both evasion and poisoning training scenarios. Additionally, this concept has been extended to encompass other widely utilized GNN models like iC-GAT and iC-SAGE. All iC-GNN models demonstrate effective defense capabilities, demonstrating comparable resilience to adversarial attacks. This underscores the superiority and scalable nature of the iC-GNN framework, providing opportunities for a variety of graph learning applications.