Detecting Unknown Shilling Attacks in Recommendation Systems.

Recommender systems are vulnerable to attacks because of their open nature. Counterfeit users give biased ratings to the items due to various objectives that may lead to the loss of user trust. The attackers use certain attack models with specific features. The existing attack detection techniques are typically attack-specific and work only when the attack features are known. They are unable to identify an unknown attack with unfamiliar features. To diminish this problem, in this paper, we propose a generalized solution that filters any attack irrespective of its design and features. We trained the classifiers with the ratings of the known authentic users using one-class SVM and PU learning models for detecting attacks, considering their ability to detect anomalies in the dataset caused by unknown attacks. The openly available MovieLens dataset has been used to assess our designed attack detection method. The experimental results show that all unknown attacks are successfully detected with 100% accuracy. The same detection accuracy is achieved for attacks with known features. [ABSTRACT FROM AUTHOR]