A Lightweight Method for Defense Graph Neural Networks Adversarial Attacks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Security

Background:

• Graph neural networks (GNNs) are increasingly used but vulnerable to adversarial attacks.
• Adversarial attacks involve subtle data perturbations that cause GNNs to produce incorrect outputs, posing significant risks.
• Existing defense strategies often require substantial computational resources and are tied to model training.

Purpose of the Study:

• To develop a defense mechanism against adversarial attacks on GNNs.
• To create a defense strategy that is computationally efficient and easy to implement.
• To maintain high defense effectiveness while reducing resource consumption.

Main Methods:

• Proposed a novel defense approach based on graph transformation.
• Implemented a lightweight and user-friendly defense strategy.
• Evaluated the method's performance against adversarial attacks on Graph Convolutional Networks (GCNs).

Main Results:

• The proposed graph transformation method demonstrated comparable defense efficacy to existing strategies, with accuracy rate returns near 80%.
• The new approach achieved this defense performance using only 10% of the runtime compared to traditional methods.
• The method proved effective in defending against adversarial attacks on GCNs.

Conclusions:

• A computationally efficient and effective defense against GNN adversarial attacks has been developed.
• Graph transformation offers a promising alternative to resource-intensive defense methods.
• The approach enhances the reliability of GNNs in real-world applications without significant overhead.