Leveraging Behavior Diversity to Detect Spammers in Online Social Networks

摘要

Online social networks have become very popular and convenient for communication. However, spammers often take control of accounts to create and propagate attacks using messages and URLs. Most existing studies to detect spammers are based on machine learning methods. Features are the key factors considered in these methods, and most documented features in existing studies can be evaded by spammers. In this study, we propose behavior features, which are based on behavior diversity when sending messages, combined with existing effective features, to build a detection system. We leverage entropy to present differences in behavior diversity between spammers and normal accounts. In the cases of evasion by periodically changing a behavior model in the sending of messages by spammers, we also introduce conditional entropy, which is calculated based on the Markov model. To achieve our goal, we have collected information from approximately 489,451 accounts including 108,168,675 corresponding messages from Sina Weibo. Through evaluation of our detection methods, the accuracy rate of this system is approximately 91.5%, and the false positive rate is approximately 3.4%.