New Metrics for Effective Detection of Shilling Attacks in Recommender Systems

Collaborative filtering techniques are successfully employed in recommender systems to assist users counter the information overload by making accurate personalized recommendations. However, such systems are shown to be at risk of attacks. Malicious users can deliberately insert biased profiles in favor/disfavor of chosen item(s). The presence of the biased profiles can violate the underlying principle of the recommender algorithm and affect the recommendations. This paper proposes two metrics namely, Rating Deviation from Mean Bias (RDMB) and Compromised Item Deviation Analysis (CIDA) for identification of malicious profiles and compromised items, respectively. A framework is developed for investigating the effectiveness of the proposed metrics. Extensive evaluation on benchmark datasets has shown that the metrics due to their high Information Gain lead to more accurate detection of shilling profiles compared to the other state of the art metrics.