Research On Adversarial Attack For Social Malicious Account Detection Model

In recent years,with the rapid growth of social users,a large number of malicious social accounts appear.The malicious account detection model based on graph classification has been widely used to detect malicious social accounts and purify the environment of social platforms.The malicious account detection model based on graph classification can be divided into graph neural network and collective classification method.The latest research shows that the graph neural network method can not resist the universal adversarial attack,resulting in a large number of node classification errors.However,the vulnerability of collective classification methods to universal adversarial attacks remains unknown.In this thesis,the universal adversarial attack of collective classification is studied,and the security risks of this method are analyzed,so as to provide a reference for designing a more robust detection model.The main work contents are as follows:(1)Propose a universal adversarial attack scheme based on universal nodes.The goal of this thesis is to make all malicious nodes establish/delete connections with universal nodes by looking for universal nodes in the graph,thereby evading detection.In this thesis,the process of finding universal nodes in graphs is regarded as an optimization problem with constraints,and the projection gradient descent algorithm is used to select the optimal universal nodes.Experiments in the Enron dataset show that94% of malicious nodes escape detection by using only seven universal nodes.(2)Propose a universal adversarial attack scheme based on injection of fake nodes.First,this thesis injects fake nodes into the graph and establishes connections with attack nodes.Then,the selection of the fake node optimal connection is regarded as an optimization problem with constraints,and the projection gradient descent algorithm is used to select the fake node optimal connection edge to ensure the attack node has the best versatility.Finally,all malicious nodes are connected to attack nodes to avoid detection.Experiments on Epinions data set show that 93% of malicious nodes evade detection by using only 10 attack nodes and 100 fake nodes in this thesis.(3)In order to verify the effectiveness of adversarial attack scheme proposed in the research work,this thesis designs and implements adversarial attack system of malicious account detection model.The system provides the above two universal attack methods and generates corresponding attack result information.Users can upload the attack result to evaluate and analyze the attack.