Defending imitating attacks in web credibility evaluation systems

Unlike traditional media such as television and newspapers, web contents are relatively easy to be published without being rigorously fact-checked. This seriously influences people's daily life if non-credible web contents are utilized for decision making. Recently, web credibility evaluation systems have emerged where web credibility is derived by aggregating ratings from the community (e.g., MyWOT). In this paper, We focus on the robustness of such systems by identifying a new type of attack scenario where an attacker imitates the behavior of trustworthy experts by copying system's credibility ratings to quickly build high reputation and then attack certain web contents. In order to defend this attack, we propose a two-stage defence algorithm. At stage 1, our algorithm applies supervised learning algorithm to predict the credibility of a web content and compare it with a user's rating to estimate whether this user is malicious or not. In case the user's maliciousness can not be determined with high confidence, the algorithm goes to stage 2 where we investigate users' past rating patterns and detect the malicious one by applying hierarchical clustering algorithm. Evaluation using real datasets demonstrates the efficacy of our approach.