| The vigorous development of deep learning in various fields has brought huge benefits and convenience to human,including image recognition,text classification,and graph data analysis.But at the same time,deep learning also has certain security risks.A lot of research work has shown that Deep Neural Networks(DNN)is vulnerable to adversarial attack,that is,a malicious attacker creates several adversarial samples with little difference from normal samples and mixes them into the target network,can make the network output wrong results.Although there is a lot of research on adversarial attack and defense in fields such as computer vision and natural language processing,it is difficult to transfer such knowledge directly to graph data.Because of the importance of graph data,the robustness of Graph Neural Networks(GNN)has become a hot topic.At present,most of the related work is to study the adversarial attack of Node Classification tasks on GNNs.The attacker only perturbs on a single graph.However,Link Prediction in Dynamic Graphs(LPDG)is an important research problem,and its applications are also diversified,such as online recommendation systems and disease infection research.Now researchers have proposed various LPDG algorithms based on graph neural networks and made progress.In this paper,we studied the vulnerability of the LPDG algorithm and proposed a blackbox evasion attack against LPDG for the first time.Our attack does not require knowledge of the specific architecture and parameters of the target model.It is a Non-target attack designed to reduce the overall effect of the target model makes it possible to get as many false predictions as possible.Considering that it is a black box attack,we regard the entire attack process as a Markov chain,combined with the Soft Actor Critic reinforcement learning algorithm,and propose an attack model called RL-based-Attack,which defines the state of the graph,the action,the reward and the stop state of the attack with the mathematical formula.In addition,we evaluated the RL-based-Attack model on three public graph datasets from different fields.The experimental results show that our method has good performance. |
