Adversarial Attacks

Structack: Structure-based Adversarial Attacks on Graph Neural Networks
Graph neural networks (GNNs) are vulnerable to adversarial attacks, but most attack methods require node attribute information. This paper introduces Structack, an uninformed attack strategy that exploits only graph structural properties without needing node attributes. By targeting links between nodes with low similarity and low centrality, Structack approaches the effectiveness of informed attacks while being computationally more efficient. The findings demonstrate that structural knowledge alone can significantly degrade GNN performance, contributing to the development of more robust graph-based machine learning methods.
Hussain Hussain, Tomislav Đuričić, Elisabeth Lex, Denis Helić, Markus Strohmaier, Roman Kern
PDF Cite Project DOI
Structack: Structure-based Adversarial Attacks on Graph Neural Networks