An AP-Based Shilling Attack Detector for Collaborative Filtering Recommendation Systems

Abstract

Collaborative filtering is a technique widely used in online recommendation systems nowadays. However, it is vulnerable from manipulation by malicious users who often create some fake account (or shilling) profiles to influence the results of recommender systems. To identify the fake users, existing algorithms usually utilize certain characteristics of shilling profiles, of which the drawbacks are the low precision and the requirement of a large size of training set. In this paper, we develop a clustering based method to find the shilling attackers by incorporating the information of user ratings and the attribute of user profiles. The users are firstly self-organizedly clustered into several groups based on the integrated information of the rating features and the attributes of user profile, then the malicious user group is identified through the GRDMA (Group Rating Deviation from Mean Agreement) values of user group. Instead of identifying attacker one by one, the proposed algorithm finds the malicious users at the collective level, which provids a novel way to analyse and detect shilling attack. The experimental results performed on MovieLens dataset demonstrate that the proposed algorithm is effective and robust in three typical kinds of shilling attack models, especially when the attack size and the filler size are sufficiently high.