Edge Gradient Exploration Attack based on Graph Neural Networks

Graph Neural Networks (GNNs) have achieved remarkable success in various graph-structured tasks due to their powerful graph representation learning capabilities. By aggregating and transforming information from the neighboring nodes, GNNs explore the structure of the graph and the features of the nodes. However, recent theoretical and empirical analyses have revealed that GNNs are particularly vulnerable to adversarial attacks. Among these, gradient-based attacks have been widely studied for their inherent efficiency and effectiveness. Nevertheless, due to the discrete nature of graph data, gradient-based attacks often result in suboptimal outcomes. Furthermore, the portability of some previously proposed attack algorithms requires further improvement. To address these two issues, we propose a novel Multi-Network Exploratory Attack (MNEA) aimed at enhancing gradient-based graph perturbations and improving their transferability. MNEA consists of two phases: the generation phase and the reconstruction phase. In the generation phase, multiple networks are utilized to generate a variety of candidate edges, preventing the generated perturbed edges from being too targeted toward a single model, which could lead to decreased transferability. The reconstruction phase introduces multiple edges for recombination, with the objective of avoiding misleading information potentially provided by the maximum gradient. Experimental results that are conducted on three publicly available benchmark datasets demonstrate that MNEA significantly outperforms baseline methods in terms of performance.