Research On Attack Detection In Recommender Systems

With the popularization and development of Internet technology,the quantity of information and the speed of the dissemination are both in explosive growth,and the emergence of recommender system has effectively alleviated the negative impact brought by the information overload problem.The collaborative filtering-based recommendation algorithm mines user preferences by analyzing historical user behavior data,thereby assisting users to discover information in which they may be interested in the future and recommending them to users.However,due to the dependency of the collaborative filtering algorithm on user information,some security issues exist in the recommender system.For example,some malicious attackers artificially inject false user information into the recommender system in an attempt to interfere with the recommendation results in order to seek benefits,which is known as “shilling attack”.The attack has seriously affected the normal working of the recommender system and also harmed the benefits of normal users.Therefore,how to deal with the shilling attack in the recommender system has become one of the hot spots in the research field of the recommender system.The problem of the detection of the shilling attack in the recommender system is studied in this thesis and it aims to achieve two goals: one is to effectively detect the shilling attack in the recommender system,and the other is to improve the efficiency of the shilling attack detection.Firstly,the working principle of the collaborative filtering recommender system,the characteristics of the “shilling attack” model are analyzed.Then,starting with the differences in scoring behaviors between the “shilling attack” users and the normal users,the user profile features are extracted for the shilling attack detection.And a user profile feature selection method Gain-User based on information gain is proposed to solve the validity problem of user profile feature,in which the feature subset obtained helps improve the accuracy of the detection algorithm effectively.Secondly,an isolated forest-based shilling attack detection method IFDM is proposed in the thesis.Compared with SVM,C4.5 and so on,this algorithm has obvious advantages in detection efficiency and maintains high detection accuracy.The parameters in the isolated forest model are analyzed and a PSO-IForest parameter optimization method based on the particle swarm algorithm to obtain the optimized parameters of the isolated forest algorithm is adapted.For isolated forests in the generation process of isolated trees,due to the problem of declining detection effect caused by the random selection of splitting features,the splitting features in a weighted manner are selected in the thesis,which prefers to select more important features,resulting that the accuracy and stability of isolated forest algorithms are improved.Finally,for the grouping characteristics of the attack,a label propagation-based attack detection model LPDM is proposed in the thesis.For the user rating behavior and the similarity between users,the label propagation weights and iteration rules are defined.This method can aggregate the user groups with similar scoring behavior under the same tag category,which can effectively discover those users with group characteristics.