Research On Online Social Network Abnormal Account Detection Algorithm

Online social networks,as the timeliest and widely used communication methods in society,have become an indispensable part of people's life.However,due to the openness of online social networks,online social networks have attracted more and more attention of attackers,which has become a new way for attackers to realize malicious behaviors.Attackers can use online social networks to post false information,pornography or even cyberbullying.These malicious behaviors not only affect the normal development of online social networks,leading to a direct decline in the user's favorability of the platform,but also serious threats to users' privacy,property safety,and even personal safety.Therefore,we conducted research on abnormal accounts in social networks,focusing on the detection of these abnormal accounts,and mainly achieved the following results:(1)Analyze and summarize the main types of abnormal accounts in online social networks,including Sybil accounts,Cyberbullying accounts,and Compromised accounts.According to the characteristics of these accounts and the existing detection methods,the characteristics such as account status,frequency of use by users,information posted by users,and the strength of connections between users are analyzed to better detect the abnormal status of the account.(2)A Sybil account detection framework based on the social network graph is proposed.Considering that the existing detection methods are likely to cause invasion of user privacy,relying heavily on assumptions and failing to work well in real networks,we propose a new detection framework.Through the combination of local classifiers and edge classifiers,not only the information in the social graph is fully utilized,but the model's dependence on assumptions is relaxed.In addition,the running effect of the random walk algorithm is stabilized by the proposed initial vertex selection method.In the end,we applied the model to three different social networks and evaluated the model through different indicators.The model showed good detection results.(3)A cyberbullying account detection scheme based on natural language processing is proposed.Considering that the existing detection models have low detection accuracy and recall rate for cyberbullying,we use two natural language analysis methods,word similarity and Fast Text,to analyze user-published content from two aspects,lexical and syntactic,in order to identify the explicit and implicit bullying meaning in the text content.Finally,we applied the model to the Formspring dataset.The accuracy rate of the model is 86.41%,and the detection precision,recall rate,and F1 value for cyberbullying text reached 59.57%,40.39%,and 48.47%,respectively.(4)A compromised account detection scheme based on user portraits is proposed.Considering the prevalence and transience of compromised accounts,we propose a lightweight detection scheme to realize real-time detection.The model mainly constructs user portraits based on the normal behavior of normal users and then monitors the account by measuring changes in account stability.In addition,our model can also update the model based on realtime feedback from users to better adapt to changes in user behavior.Finally,we conducted experiments on a dataset consisting of 40 Twitter accounts.The model's false-positive and false-negative rates are 0.725% and 5.71%,respectively.