Graph Based Virtual Node Neural Network Malicious Account Identification In Login Scenarios

With the rapid development of the mobile Internet,a large number of Internet companies have emerged.In order to expand the influence of their products,grab highquality users,and seize target traffic,these companies spend a lot of marketing funds on APP products every year.And these huge marketing funds are also being coveted by unscrupulous people.These industrialized and large-scale black products illegally obtain marketing funds through illegal means such as fake accounts and batch registration,which has brought incalculable losses to enterprises,and society.For Internet companies,preventing black production risks has become an urgent issue.Many companies have set up risk control departments internally,and use various algorithms and rules to compete with black production technology.The login and registration scenario is a necessary process for each user to log in to the APP,so prevention and control in this scenario is particularly important.There are also many existing algorithms in this scenario: rule-based methods,supervised methods,deep learning methods for user behavior,etc.,but these algorithms have their limitations in this scenario.The application of graph algorithms in this scenario has its advantages,but it is also necessary to improve and innovate the graph neural network based on the characteristics and difficulties of the login scenario.Therefore,this paper explores the algorithm application of graph neural network in the login and registration scenario of an APP product.First,through data analysis and expert experience,find out the characteristics of the malicious behavior of the black people in the login scenario,and use it as the connection method on the graph;secondly,this connection characteristic makes the users form an isolated community,which is not conducive to the parameter characteristics.Passing on the graph,this paper also innovatively proposes community virtual nodes and clustering edges to increase the connectivity and training efficiency of nodes in the graph.Finally,the model is trained for prediction and community mining.And compare the results of the new algorithm with the baseline algorithm to explore the improvement.Finally,the algorithm can achieve an accuracy rate of 0.968 and an AUC of 0.783,which is an improvement over the baseline algorithm.This can also show that while ensuring high accuracy,it can also help risk control business personnel to monitor business operations and provide algorithmic decision-making basis for abnormal user interception.