Leveraging Temporal Pattern To Detect Malicious Accounts In Online Social Network

Online social networks,such as Facebook and Sina Weibo,have become the most popular platforms for information sharing and social activities.More and more malicious users have utilized social networks as a new way to spread spam information using malicious accounts.Due to the harmful effects of malicious accounts on the social networks,many methods are proposed to solve this problem.The existing methods are divided into two types,which are based on machine learning and graph-based algorithm.The methods based on machine learning extract some features from the accounts' profiles,behaviors,and tweet contents,and then build supervised learning classifiers.The graph-based methods start with the graph structure of online social networks,and propose many iterative graph algorithms that are based on Page-rank to analyze the accounts in social networks,and then identify the malicious accounts according to the results.However,as time goes on,malicious users are upgrading their behavioral strategies rapidly to race with the development of detection systems.For the two types of detecting methods mentioned earlier,malicious users have found some ways which are very mature to avoid the detection based on these methods.As a result,researchers have proposed some new approaches to meet these challenges.However,these new methods still do not solve the most critical problem.whether an individual account is malicious is inferred by these methods based on its characteristics at a single instant of time.In this study,we introduce temporal factors into the detection of malicious accounts by inspecting the activities of accounts over an extended period of time and offer a detecting framework to identify the spammers that evade detection by changing their strategies.To achieve our research goals,we collect the profiles of a vast number of social network accounts and track their activities over a series of points of time.A window-based dynamic model is used to assess the temporal evolution patterns of accounts and uncover a clear distinction between legitimate and malicious users concerning different aspects of the temporal patterns.Based on the dynamic model,new temporal user features are designed to detect malicious accounts.Instead of using these features to identify malicious accounts directly,we investigate the similarity in the temporal patterns of different malicious accounts,and conduct a clustering algorithm on accounts by abstracting their dynamic changes into feature vectors.The results indicate that it is relatively easier to group malicious accounts into the same clusters.We combine the new features with the clustering results to build a machine learning classifier for accurate detection of malicious accounts.Finally,we evaluate our detecting framework using the real-world dataset and demonstrate the effectiveness of our approach by comparing it with two conventional spammer detection methods.