Strong Target Attack on Hypergraph Neural Networks via Label Poisoning and Structure Modification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Hypergraph Neural Networks (HGNNs) excel at modeling complex relationships.
• HGNNs exhibit adversarial vulnerabilities, posing security risks, especially in targeted attacks.
• Existing attacks on HGNNs lack precise control over the misclassification target class.

Purpose of the Study:

• To propose a novel framework for strong target attacks against HGNNs.
• To address the research gap in precisely controlled misclassification attacks.
• To enhance the security analysis of HGNNs under sophisticated adversarial threats.

Main Methods:

• Developed the Strong Target Attack framework for HGNNs (STALS) using label poisoning and structure modification.
• Implemented feature similarity and hypergraph structure adaptability for optimal target class selection.
• Utilized a gradient-guided greedy hyperedge reconstruction strategy for efficient mislabeled information propagation.

Main Results:

• STALS demonstrated excellent performance in achieving directed misclassification.
• The framework successfully misclassified source-class nodes into predefined target classes.
• STALS significantly outperformed existing baseline methods in success classification rate across four datasets.

Conclusions:

• STALS effectively fills the research gap for strong target attacks on HGNNs.
• The proposed method offers a robust approach to evaluating HGNN security.
• Further research into robust HGNN defenses against such targeted attacks is warranted.