FHA - Fast Heuristic Attack Against Graph Convolutional Networks.

Recent years have witnessed a significant growth in Graph Convolutional Networks (GCNs). Being widely applied in a number of tasks, the safety issues of GCNs have draw the attention of many researchers. Recent studies have demonstrated that GCNs are vulnerable to adversarial attacks such that they are easily fooled by deliberate perturbations and a number of attacking methods have been proposed. However, state-of-the-art methods, which incorporate meta learning techniques, suffer from high computational costs. On the other hand, heuristic methods, which excel in efficiency, are in lack of satisfactory attacking performance. In order to solve this problem, it is supposed to find the patterns of gradient-based attacks to improve the performance of heuristic algorithms. In this paper, we take advantage of several patterns discovered in untargeted attacks to propose a novel heuristic strategy to attack GCNs via creating viscous edges. We introduce the Fast Heuristic Attack (FHA) algorithm, which deliberately links training nodes to nodes of different classes. Instead of linking nodes fully randomly, the algorithm picks a batch of training nodes, which are of the same type, and links them to another class each time. Experimental studies show that our proposed method is able to achieve competitive attacking performance when attacking against various GCN models while significantly outperforming Mettack, which is the state-of-the-art untargeted structure attack, in terms of runtime.