Securing Collaborative Filtering Against Malicious Attacks Through Anomaly Detection

Collaborative filtering recommenders are highly vulner-able to malicious attacks designed to affect predicted ratings. Previous work related to detecting such at-tacks has focused on detecting profiles. Approaches based on profile classification to a large extent depend on profiles conforming to known attack models. In this paper we examine approaches for detecting sus-picious rating trends based on statistical anomaly de-tection. We empirically show these techniques can be highly successful at detecting items under attack and time intervals when an attack occurred. In addition we explore the effects of rating distribution on detec-tion performance and show that this varies based on distribution characteristics when these techniques are used