AN-GCN: An Anonymous Graph Convolutional Network Against Edge-Perturbing Attacks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph convolutional networks (GCNs) are susceptible to adversarial attacks that manipulate graph structures.
• Existing defenses against edge-perturbing attacks on GCNs lack theoretical grounding and robust solutions.

Purpose of the Study:

• To theoretically prove the vulnerability of GCNs to edge-perturbing attacks in node classification.
• To propose a novel defense mechanism, the anonymous GCN (AN-GCN), against these attacks.

Main Methods:

• Generalized the formulation of edge-perturbing attacks and provided a strict theoretical proof of GCN vulnerability.
• Introduced a node localization theorem to explain GCN training dynamics.
• Developed a staggered Gaussian noise-based node position generator and a spectral graph convolution-based discriminator.
• Proposed an optimization method for the generator and discriminator components.

Main Results:

• Validated the effectiveness of the general edge-perturbing attack (G-EPA) model in manipulating node classification.
• Demonstrated that AN-GCN achieves 82.7% node classification accuracy without using edge information.
• Showcased AN-GCN's security against edge-perturbing attacks by rendering edge information irrelevant to attackers.

Conclusions:

• GCNs are theoretically vulnerable to edge-perturbing attacks, necessitating robust defense strategies.
• AN-GCN offers a secure and effective solution for node classification tasks, outperforming existing GCN models under attack scenarios.