Using Trust and Reputation for Detecting Groups of Colluded Agents in Social Networks

One of the most common types of malicious behavior in social networks is represented by collusion, which consists of a secret cooperation between two or more agents providing mutual, highly positive feedback to each other. This collusion creates misleading advantages for the involved agents, deceiving others and distorting the actual reputation perception of the colluding members. Although the well-known EigenTrust algorithm can be fruitfully used to detect colluded agents, two important issues arise which limit its effectiveness: (i) it requires input information about which agents can be a-priori considered particularly trustworthy, and (ii) it is not designed to handle situations in which we have several, different groups of colluded agents. These problems lead EigenTrust, to produce a significant number of false positives in some real situations. In this paper, we address the aforementioned issues. We introduce an automatic procedure to provide EigenTrust with the necessary inputs, and we propose an appropriate algorithm that combines EigenTrust with a clustering process. This procedure groups agents based on their reputation scores to tackle the presence of different groups of colluded agents. Through experiments, we demonstrate that our method, while maintaining the same effectiveness as EigenTrust in detecting malicious agents, is significantly more capable of avoiding the generation of false positives.