Graph Structure Adversarial Attack Design Based on Graph Attention Networks

Graph Neural Networks (GNNs) have been demonstrated to be effective as node classifiers under ideal conditions. However, minor perturbations in the graph may significantly affect the classification performance of GNNs. Therefore, it is imperative to investigate potential attack methods against GNNs, so that more robust networks or effective defense models could be developed. In this paper, a novel inconspicuous adversarial attack model, termed Graph Attention Adversarial Attacks (GATTACK), is proposed. GATTACK leverages Graph Attention Networks (GATs) to acquire the surrogate model, employs structure attacks targeted at specific nodes with constraints to ensure imperceptible perturbations, and takes full advantage of GAT’s sensitivity to the topology of graphs. The efficiency of GATTACK is validated through comparative experiments, highlighting its performance and utility.

1 Introduction

Many types of real-world data can be represented by topological graphs with certain features and structures, such as biomolecules [1], social networks [2], transportation networks [3], Internet of Things (IoT) networks [4, 5] and so on. Typically, the entities and their attributes in these datasets are treated as points with features in the topologies, while the interactions between the entities are abstracted as interconnected edges. By transforming data into topological graphs, Graph Neural Networks (GNNs) can be employed to address practical problems, sparking a surge of related research in recent years.