GAA-PPO: A novel graph adversarial attack method by incorporating proximal policy optimization

The Graph Convolutional Network (GCN) has demonstrated impressive performance in processing graph structured data. However recent studies have revealed that GCN is vulnerable to adversarial attacks, where a small amount of data modification can significantly affect the performance of the GCN models. While most existing studies node injection attacks with graph reinforcement learning by considering gradient information, they still suffer from the problems that the step size of the policy gradient is difficult to determine, and the attack effect needs to be further improved. In light of the above issues, this paper proposes a Graph Adversarial Attack method by incorporating Proximal Policy Optimization named GAA-PPO, which fills subtasks of sequentially generating features and links for injected nodes without modifying existing nodes or edges. GAA-PPO comprises two main components: node injection attack network (actor network) and value prediction network (critic network). Specifically, the actor network leverages a node generator and an edge sampler to generate appropriate features and edges for the injected nodes. Notably, a novel edge sampler that incorporates Approximation Personalized Propagation of Neural Prediction (APPNP) is introduced to effectively propagate malicious features of the injected nodes. On the other hand, the critic network evaluates the performance of the perturbed graph at each stage. To enhance the stability of the algorithm, GAA-PPO employs the importance sampling technique of Proximal Policy Optimization (PPO) during the training process. Extensive experiments on three publicly benchmark datasets show that GAA-PPO yields significant performance advantages over the state-of-the-art method.