Research On Power User Attack Detection Approach In Recommender System

Recommender systems have become an effective tool of information retrieval, they use collaborative filtering technology through collecting users’ preference information and filtering out the information from massive amounts of information to satisfy the users’ interests and recommend the information to users actively, which solves the problem of information overload to a certain extent. Due to their natural openness and dependence on users’ preference information, however, lead the system to be vulnerable to attack by malicious users. Attackers may manipulate the recommendation results to their benefits by injecting a lot of false information into recommender systems. To ensure the security of the recommender systems, domestic and foreign researchers have proposed a variety of attack detection methods. However, attackers are constantly looking for new ways to mount the attack too. On the basis of analysis for current research in this area, this paper has conducted deep research on novel Power User Attack. In a Collaborative Filtering Recommender System context, power users are those who can exert considerable influence over the recommendation outcomes presented to other users.Firstly, aiming at the problem that the existing power user selection methods can not effectively identify the influential power users, from the perspective of social network analysis, an aggregate neighbor degree centrality and neighbor similarity method based on social network degree centrality concepts is proposed. Through considering the presence or absence of relationship between users and the strength of relationship between users can accurately identify the influential power users in the system.Then, the existing recommender system attack research has examined similarity-focused statistical models of user ratings, which has a fixed generation mode. This attack can easily be detected by some attack defense technology. This paper envision attacker shift to power users based explicitly on measures of influence to mount attack, thus affecting the recommender system predictions and top-N recommendation lists. On the basis of the selected real power users, to build power user attack model, and simulate power user attack profiles.After that, aiming at the problem that the existing attack detection algorithms can not accurately detect novel power user attack, an unsupervised power user attack detection algorithms is proposed. From the perspective of rating features of power users with aggressive behavior, the algorithm first find the attacked item, and then detect the power user attack profiles according to the attacked item.Finally, the comparative experiments are conducted with the existing power user selection methods and attack detection algorithms on the Movie Lens dataset.