Research On Shilling Attack Detection Algorithm For Collaborative Filtering Systems

Collaborative filtering system largely alleviates the problem of "information overload" on the network.However,due to the vulnerability of the system,the malicious users inject biased profile into such a system and make themselves become the neighbor of normal users so as to manipulate the system and achieve the purpose of commercial competition.This behavior leads to the deviation of the recommended results,a decline in recommendation services,a lack of user confidence in the recommendation system,so how to detect the shilling attack users become an urgent problem to be solved.In order to control the impact of recommended attacks,domestic and foreign research scholars use a lot of methods for detecting,the core of the problem is how to improve the detection performance.In this paper,we will make a deep research on the shilling attack detection algorithm for collaborative filtering system based on the two design ideas of supervised learning and unsupervised learning.Firstly,the existing supervised detection algorithms have a low precision when they detect shilling attack,we propose an shilling attack detection algorithm based on item popularity and novelty degree features.This algorithm based on the differences of the way users choose items to rate between genuine and attack profile,we extract several features that effectively distinguish normal users and shilling attack users from the angle of popularity and novelty;On this basis,we propose an ensemble detection framework and this process is that through the Boosting technology to produce a large number of different base classifiers,and then combining with the weight of the base classifier to obtain the detection result.Secondly,the existing unsupervised detection algorithms have a low precision when they detect shilling attack.we propose an shilling attack detection algorithm based on multiple features hybrid decision.The algorithm selects multiple features using K-means and threshold judgment the two kinds of detection algorithms to determine user clusters,and using the users support to unify the user cluster.On this basis,using collaborative rating deviation to determine the attack item,and then filter misjudgment of the realprofile to obtain the detection results.Finally,the corresponding experiments are designed for the two detection algorithms,and the results of the comparative experiments are analyzed to verify the effectiveness of the proposed algorithm.