Spectral adversarial attack on graph via node injection.

Area of Science:

• Graph Neural Networks (GNNs)
• Adversarial Attacks
• Network Security

Background:

• Graph Neural Networks (GNNs) are widely used but vulnerable to adversarial perturbations.
• Existing node injection attacks are limited to local graph structures.
• This limits their effectiveness in compromising GNN performance.

Purpose of the Study:

• To propose a novel global graph attack method, Spectral Node Injection Attack (SpNIA).
• To enhance the effectiveness of adversarial attacks on GNNs by considering global graph properties.
• To improve upon existing node injection attack limitations.

Main Methods:

• SpNIA leverages spectral distance by maximizing the Euclidean distance of eigenvalues from Laplacian matrices.
• A novel optimization framework is developed to handle differing matrix dimensions and interconnected injected nodes.
• Gradient-based methods are employed to solve the optimization problem.

Main Results:

• Experiments demonstrate a significant decrease in GNN performance on benchmark datasets.
• SpNIA effectively leverages limited adversarial budgets through spectral distance.
• The proposed method shows feasibility and effectiveness in compromising GNNs.

Conclusions:

• SpNIA represents a significant advancement in global graph attack methodologies.
• The spectral distance approach offers a more potent strategy for adversarial attacks on GNNs.
• This research highlights the need for more robust GNN defenses against sophisticated attacks.