Detection of Profile-injection attacks in Recommender Systems using Outlier Analysis

摘要

E-Commerce recommender systems are vulnerable to different types of profile-injection attacks where a number of fake user profiles are inserted into the system to influence the recommendations made to the users. In this paper, we have proposed three strategies of detecting such attacks with the help of outlier analysis. In all these strategies, the attack-profiles are considered as outliers in the user rating dataset. Firstly, we have used Partition around Medoid (PAM) clustering algorithm in detecting the attack-profiles. An incremental version of the PAM algorithm has been applied and tested for evaluating the performance of the system in identifying the attack profiles when they come into the system. Experiments show that though PAM is able to detect attack profiles with larger number of filler items very well, a percentage of attack profiles with smaller number of filler items is not included in outlier clusters-they are included in large clusters. Secondly, we have applied a PAM-based outlier detection algorithm to find these attack profiles in large clusters. Finally, an angle based outlier detection strategy is used for finding attack profiles in the database under attack.