Shilling attack detection in Collaborative Recommender Systems using a Meta Learning strategy

Collaborative Recommender Systems suggest items to a user based on other users past behaviour (items they once bought, viewed or selected and/or ratings they gave to those items). They are very effective in generating meaningful recommendations to a group of users for products or items that might interest them. However, since Collaborative filtering techniques depend on outside sources of information they are susceptible to profile injection attacks popularly known as shilling attacks. Shilling is a process in which syndicating users can connive to promote or demote a certain item. These mischievous users can consciously inject shilling profiles in an effort to bias the recommender system to their advantage. In this paper we seek to understand the degree to which shilling attacks can harm recommender systems and how these attacks can be detected. Firstly, we evaluate the vulnerabilities of collaborative filtering techniques in providing reliable recommendations. We study various attack strategies that manipulators use to attack recommender systems. Secondly we investigate the most suitable features that can be used to adequately identify shilling attacks. We propose the combiner strategy that combines multiple classifiers in an effort to detect shilling attacks. The diversity measure is used to determine the most suitable combination of classifiers. In this paper, we made use k-Nearest Neighbour, Support Vector Machines and Bayesian Networks as the initial base classifiers. The Naïve Bayes was used as a Meta Classifier. The proposed Meta-Learning classifier gave an overall performance of 99% and was found to be more superior to Neural Networks and k-Nearest Neighbor.