Multi-view contrastive learning for graph adversarial defense.

Area of Science:

• Graph Neural Networks
• Adversarial Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) are powerful for graph tasks but vulnerable to adversarial attacks.
• Existing defenses often focus on single views or limited information, neglecting critical local and global graph insights.
• Developing robust GNN frameworks is a significant research challenge.

Purpose of the Study:

• To propose a novel defense mechanism, COLA (Contrastive Learning for Graph Adversarial Defense), to enhance GNN robustness.
• To address the limitations of current defense methods by incorporating both local and global graph information.
• To improve the reliability of node representations against adversarial perturbations.

Main Methods:

• Generated two augmented graph views using edge directionality and graph diffusion, incorporating structure, features, and supervised information.
• Employed multi-view contrastive learning to encode local and global graph information via distinct contrast paths.
• Constructed different contrast paths to derive robust node representations.

Main Results:

• Validated COLA on seven diverse benchmark datasets (four homophilic, three heterophilic).
• Demonstrated COLA's effectiveness in resisting various adversarial attacks.
• Achieved superior performance compared to state-of-the-art baseline methods.

Conclusions:

• COLA significantly improves the robustness of GNNs against adversarial attacks.
• The proposed multi-view contrastive learning approach effectively leverages local and global graph information.
• COLA offers a promising direction for building more resilient GNN models.