Shilling attack detection in collaborative filtering recommender system by PCA detection and perturbation

Collaborative Filtering (CF) recommender systems have been widely used in many applications due to its satisfying performance in personalized recommendation. Recent studies show that a CF recommender system is vulnerable to shilling attacks in which shilling profiles are injected into a system by an adversary. Many attack detection methods have been proposed to defense against shilling attacks. Unsupervised Primary Component Analysis (PCA) is one of the most effective detection methods. However, its efficacy relies on the prior information, i.e. the number of shilling profiles in a recommender system. In this paper, an un-supervised shilling attack detection method which combines PCA and perturbation is proposed. In the proposed method, PCA is applied before and after inserting Gaussian noise to each user profile. Shilling attacks are detected by combining the results of the two PCA. Compared with using PCA alone, the proposed method achieves higher accuracy in experiments. This indicates that injecting perturbation to samples is helpful in shilling attack detection.