An Adversarial Attack Defense Method Based on Graph Purification

With recent developments, graph neural networks (GNNs) have shown great potential in a variety of graph-related tasks, and their applications have received a lot of attention. However, current research has shown that graph neural networks are vulnerable to adversarial attacks, a problem that hinders their deployment in real-world tasks. In order to solve this problem, a defense method based on graph purification is proposed in this paper. Specifically, the method consists of three different modules that together implement adversarial defense. Firstly, a special strategy is used in the graph feature enhancement module to train the graph feature enhancement model, so that the model can generate robust graph features. In addition, the graph structure enhancement module is used in this method. First, the similarity coefficients of each node and its neighbors are obtained by the method of node feature similarity evaluation, and the low-rank graph is generated by SVD method. Then, the two generated graphs are adaptively fused to generate the final enhanced graph. Finally, the graph-adversarial defense problem is modeled as a Markov decision process (MDP), and then reinforcement learning is used to solve the optimal defense strategy. Experiments on multiple datasets show that our model outperforms current state-of-the-art methods.