Adaptive Adversarial Attack on Graph Embedding via GAN

Abstract

Graph embedding method learns the low-dimensional representation of graph data, which facilitates downstream graph analysis tasks, such as node classification, graph classification, link prediction and community detection. With the in-depth study of graph analysis tasks, the issues of excessive data mining by graph embedding methods have become increasingly prominent, a number of graph embedding attack methods have been put forward. Inspired by promising performance of generative adversarial network, this paper proposes an adaptive graph adversarial attack framework based on generative adversarial network (AGA-GAN). We use the game between a generator and two discriminators with different functions to iteratively generate the adversarial graph. Specifically, AGA-GAN generates the adversarial subgraph according to different attack strategies to rewire the corresponding parts in the original graph, and finally form the whole adversarial graph. To address the scalability problem of existing graph embedding attack methods, we consider the adaptively selected K-hop neighbor subgraph as the attack target instead of the original graph. Experimental study on real graph datasets verifies that the AGA-GAN can achieve state-of-the-art attack performance in most node classifications.