GHAttack: Generative Adversarial Attacks on Heterogeneous Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Heterogeneous Graph Neural Networks (HGNNs) are increasingly used but vulnerable to adversarial attacks.
• Current HGNN attack methods are computationally expensive, limiting their use during inference.

Purpose of the Study:

• To develop an efficient and effective adversarial attack method for HGNNs.
• To address the computational inefficiency of existing HGNN attack strategies.

Main Methods:

• Introduced Generative Heterogeneous Attack (GHAttack), a novel generative attack approach.
• Developed a perturbation generator trained via an optimization problem, utilizing an HGNN backbone and relation-aware output layer.
• Enabled perturbations to modify edges within heterogeneous graph relations for enhanced attack effectiveness.

Main Results:

• GHAttack demonstrated high efficiency and excellent effectiveness in experiments.
• Validated across ten representative HGNNs and six datasets.
• The generative approach allows for rapid perturbation generation through a simple forward pass.

Conclusions:

• GHAttack offers a computationally efficient solution for adversarial attacks on HGNNs.
• The method is effective in degrading HGNN performance by perturbing graph structures.
• This work advances the field of adversarial robustness for graph-based machine learning models.