Abstract

Title: Evaluation of Profile Injection Attacks In Collaborative Recommender Systems
Published: May 2006
Authors: Chad Williams, Runa Bhaumik, JJ Sandvig, Bamshad Mobasher, Robin Burke
Abstract: Significant vulnerabilities have been identified in collaborative recommender systems. The open nature of collaborative filtering allows attackers to inject biased profile data and force the system to ``adapt'' in a manner advantageous to them. Previous work has shown both user-based and item-based recommender systems are vulnerable to the segment attack model. In this paper we focus on two techniques that may be used to reduce the impact of the segment attack and also examine their robustness against traditional attack models. One technique is a model-based algorithm based on probabilistic latent semantic analysis that offers significant improvements in stability and robustness against all identified attack profiles. Second, we analyze the effectiveness of a supervised classification approach to detection we have introduced for protecting systems against traditional attacks.
Keywords: Attack Models, Bias Profile Injection, Collaborative Filtering, Recommender Systems, Shilling, Pattern Recognition
Full Paper: [pdf]