DGHSA: derivative graph-based hypergraph structure attack.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Hypergraph Neural Networks (HGNNs) excel at higher-order tasks but are susceptible to adversarial attacks.
• Existing attacks on HGNNs primarily focus on targeted scenarios, lacking practical untargeted strategies.

Purpose of the Study:

• To propose a novel derivative graph-based hypergraph attack (DGHSA) for untargeted adversarial scenarios.
• To address the vulnerability of HGNNs to attacks that degrade global performance.

Main Methods:

• DGHSA employs a two-stage approach: candidate set generation and evaluation.
• Gradients from trained HGNNs inform hypergraph structure modifications for attack generation.
• A derivative graph metric assesses node similarity to identify optimal perturbations.

Main Results:

• Extensive experiments on four datasets demonstrate DGHSA's effectiveness.
• The proposed attack significantly degrades HGNN performance in node classification tasks.
• DGHSA successfully reduces the global performance of HGNNs.

Conclusions:

• DGHSA presents a practical and effective method for untargeted adversarial attacks on HGNNs.
• The derivative graph-based approach offers a new perspective for analyzing HGNN robustness.
• Further research can explore defenses against such untargeted hypergraph attacks.