A New Approach to Detect User Collusion Behavior in Online QA System

Sybils and Sybil attacks are problems born with social networks. In online QA application Afanti, Sybil users collude with each other to mimic normal users and lower the possibility to be detected. In this paper, we stated this phenomenon and put forward a new approach that can detect user collusion behavior which cannot be detected before. In this approach, we defined interaction weight between users to describe the collusion, clustered users by this weight and labeled users as Sybil or normal through clustering coefficient. We also proposed a plan for deploying this approach in large-scale system and pointed out the key part and how to improve it. The experiment shows the accuracy of our approach is 93.5% in detecting Sybil questioners and 97.4% in Sybil answerers. Our approach can also recall many Sybils which cannot be detected by the original detection system.