Cluster-guided adversarial graph contrastive learning.

Abstract

Graph contrastive learning (GCL) has shown its promising ability in unsupervised graph representation learning. Recent studies have shown that GCL-based methods are vulnerable to adversarial attacks, where small perturbations to graph structures and node features can mislead the models and limit their deployment in security-critical domains. We revisit the robustness of graph contrastive learning against adversarial attacks and identify two factors that affect model robustness. First, existing methods mainly rely on single-level contrastive learning, which focuses on local neighborhood structures and may overfit structural noise in adversarial scenarios. Second, they often overlook the reliability of sample pairs, although selecting reliable positive and negative samples is crucial to representation quality, especially when the graph is under attack. To address these issues, we propose a Cluster-guided Adversarial Graph Contrastive Learning (CAGCL) approach, which combines dual-level contrastive learning with cross-view mapping of high-confidence samples. With category-level guidance derived from reliable samples, CAGCL strengthens the interaction between adversarial and augmented graph views and learns more robust and discriminative representations. Experiments on multiple benchmark datasets under untargeted and targeted attacks demonstrate that the proposed approach outperforms state-of-the-art baselines.