MCGCL:Adversarial attack on graph contrastive learning based on momentum gradient candidates

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Adversarial attacks on graph contrastive learning face challenges due to discrete graph structures, leading to unreliable gradients and local optima.
• Existing methods struggle with gradient accuracy and convergence speed in adversarial graph attack scenarios.

Purpose of the Study:

• To propose a novel adversarial attack method using momentum gradient candidates for unsupervised graph contrastive learning.
• To enhance the reliability of structural gradients and overcome the local optima problem in adversarial attacks.
• To improve the convergence speed and success rate of adversarial attacks in graph representation learning.

Main Methods:

• Transforming back-propagated gradients into momentum gradients by incorporating previous gradient information.
• Guiding gradient updates with momentum gradients to accelerate convergence and improve accuracy.
• Ranking candidate adversarial samples based on saliency derived from summed momentum gradients across two views.

Main Results:

• The proposed momentum gradient candidate method significantly improves convergence speed compared to existing adversarial attack strategies.
• Experimental results on three datasets show superior performance in link prediction tasks, outperforming supervised baselines.
• The adversarial attack method demonstrates strong transferability across different graph representation models.

Conclusions:

• The momentum gradient candidate approach effectively addresses the limitations of discrete graph structures in adversarial attacks.
• This method offers a more robust and efficient strategy for generating adversarial samples in graph contrastive learning.
• The demonstrated transferability highlights the broad applicability of the proposed attack technique in graph representation learning.