Research On User Profile Attack Detection Algorithm For Collaborative Filtering Recommender Systems

Nowadays, there are serious security problems existing in the personalized collaborative filtering recommender systems. The malicious users can make the recommendations biased through introducing many fake user profiles to the recommender systems, which reduce the confidence of users to recommender systems. So, how to ensure the security of personalized collaborative filtering recommender systems has become an urgent problem. On the basis of comprehensive analysis for the current research in this area, this paper has further deep research on the security problems of collaborative filtering recommender systems.Firstly, this paper analyses the characteristics of attack profiles and the corresponding attack detection algorithms. Aims at some problems of existing detection algorithms, considers the purpose of attacks, an approach to compute item rating deviation is proposed through analyzing the different characteristics between target item and filler items of attack profiles. Then recognizing the target item through exploiting the item rating deviation, on the basis of that, an user profile injection attack detection algorithm based on the target item identification is proposed, it overcome some defects of existing algorithms and improve the detection accuracy.Secondly, this paper considers from the perspective of items, and detects the suspicious trends of item rating distribution caused by shilling attacks. Aims at the defects and drawbacks of existing anomaly detection algorithms, analyzes the historical rating distribution characteristics of item-self according to the time series. And then the confidence interval of rating feature for each item is calculated, it is used to make sure the item whether is under attack through monitoring the rating behavior during the new coming period of time. Finally, this paper gives the experimental evaluations of the algorithms proposed in this paper, compares and analyzes the experimental results with existing algorithms, and also makes the conclusions and prospects for the further research.