Defense Method Against Shilling Attack In Collaborative Filtering

Making a suitable recommendation to users is a challengeable problem in the era of big data,Collaborative filtering is one of the most popular and efficient solutions of this problem due to its convenience and high efficiency.“Collaborative” means that users collaborate together in order to provide useful information in some sense.However,an adversary may generate malicious users to mislead decisions of a recommendation system on purpose in reality.Shilling attack which injects fake user profiles is one of popular attacks.The attack not only changes the recommendation frequency of an item,but also reduce trust of the system.As a result,it is important to defend shilling attack,and increase security of a collaborative filtering recommender system.The existing defence methods of shilling attack may have poor performance when facing the attack based on item correlation.This is because the rated items of real users are correlated,but most of attacks do not consider the correlation in their strategies.This study aims to analyze the existing defense methods against shilling attack in collaborative filtering.We firstly analyzed difference between item correlation of real users and the one of fake users.An attack method and also a defense method based on item correlation are then introduced..The major contributions of this study: according to the difference between item correlation of real users and malicious users is identified,an shilling attack method is proposed.Experimental results show that our attack method misleads a recommendation system efficiently and also evade the existing attack detection.It indicates the weakness of the existing attack detection methods.On the other hand,a new detection method based on item correlation is proposed to enhance the safety of a recommender system.By measuring the item correlation of user profiles,the influence of shilling attack on a recommender system is reduced.The propose detection method also detect contaminated by different kinds of attack efficient in the experiment.The results confirm that the proposed method is robust and efficient.