ERMAV: Efficient and Robust Graph Contrastive Learning via Multiadversarial Views Training

Area of Science:

• Graph Representation Learning
• Machine Learning Security
• Artificial Intelligence

Background:

• Graph contrastive learning (GCL) is crucial for graph representation learning.
• Existing GCL methods are vulnerable to adversarial attacks.
• Current robust GCL approaches are computationally expensive and lack scalability.

Purpose of the Study:

• To propose an efficient and robust GCL framework against adversarial attacks.
• To address the inefficiency and scalability issues of existing robust GCL methods.

Main Methods:

• Introduced ERMAV (efficient and robust GCL via multi-adversarial views training).
• Generated adversarial views by attacking node attributes and latent representations on subgraphs.
• Employed efficient attack methods for dynamic adversarial perturbation generation.

Main Results:

• ERMAV outperforms state-of-the-art GCL methods on original graphs.
• ERMAV demonstrates superior robustness compared to existing methods on attacked graphs.
• Extensive experiments on seven real-world datasets validate the framework's effectiveness.

Conclusions:

• ERMAV offers an efficient and scalable solution for robust GCL.
• The proposed multi-adversarial views training enhances GCL resilience.
• ERMAV shows significant potential for real-world applications requiring robust graph representations.