Detection of Shilling Attacks on Collaborative Filtering Recommender Systems by Combining Multiple Random Forest Models

Collaborative filtering recommender systems are one of the essential recommender systems and are still widely used in combination with other algorithms to make predictions for users. However, they are vulnerable to shilling attacks, and if there isn’t any detection system to prevent those attacks, original recommendations can be heavily influenced to benefit the attackers. Designing attack-resistant recommendation systems is not an easy task, and many researchers have tried to tackle that problem. In this paper, a new approach that combines multiple random forest models is proposed. Each of the random forest models is specialized in detecting one group of shilling attacks, and then all the models are combined into an ensemble model. Experimental results show that the proposed ensemble is capable of detecting attack profiles at high rate without causing significant bias in the original recommendation system.