Research On Shilling Attack Detection Based On Target Item Analysis In Recommender Systems

With the rapid development of computer science and communication technologies and the explosive growth of information resources, how to find useful information from the mass of information has become a key issue to be solved. As an important tool to alleviate the information overload, recommender systems push information to users actively, which have been widely used. However, due to the open nature of recommender systems, attacks against recommender systems continue to occur. Malicious users deliberately falsified user profiles by injecting into recommender systems, trying to change the rankings of target items in the system, which we call this behavior shilling attack. In order to reduce and eliminate the impact of the attacks on the recommender systems, attack detection techniques have aroused widespread concern in academia, and have become hot topic of the current research in recommender systems.Existing shilling attack detection methods usually extract user profiles’ eigenvalues, and detect shilling attacks based on extracted user profiles’ eigenvalues, and have achieved some results. However, there are some problems in exsiting shilling detection method. Existing methods do not take full advantage of the group characteristics of shilling attacks; can not effectively describe the known types of attacks in recommender system; can not effectively detect unknown type of recommender system attacks, as well as the efficiency is not high with data increases. In this thesis, considering the defects and deficiencies in exsiting shiling attack detection methods, on the basis of existing profile attribute extraction technology, profile features with a better description capabilities and high detection performance attack detection method are proposed.The main contents of the thesis are as follows:① Overview on recommender system and attack detection in recommender systems. Concepts of recommender systems, research status and major difficulties and challenges are introduced; recommendation system similarity calculation methods are analyzed; evaluation metrics and some profile features are introduced; recommender system attribute extraction techniques arereviewed.② Considering group characteristics of shilling attacks and the sparsity of rating matrix, a shilling detection framework based on target item analysis is proposed. Firstly, user profiles with high possibility of being attack profiles are found; secondly, a rating matrix from these suspected attacks is constructed; thirdly, intentions and target items using shilling detection framework are analyzed; attack profiles are finnally retieved in the end.③ By analyzing the distribution of features of profiles in real profiles and attack prifiles, two attack detection methods, RD-TIA and DeR-TIA are proposed based on target item analysis method. RD-TIA algorithm is used to detect the attacks of average attacks model and random attack model. Considering RD-TIA can not effectively detect attacks bandwagon attack model and segemnet attack model. A new attack detection algorithm DeR-TIA is proposed, which can detect attacks in random attack model, average attack model, bandwagon attack model, segment attack model and combined attack models consisting the four attack models. The experimental results show that: Compared with other classical unsupervised detection method, RD-TIA has higher accuracy and low false positive rate in detecting random attack model and average attack model; DeR-TIA algorithm has strong adaptability, the model is able to detect a variety of attack models.④ Detection accuracy of attack detection algorithm based on unsupervised is low when there is little priori knowledge of the recommender system. On the other side, class imbalance problem exists in attack detection algorithms based on supervised mehtod. A new attack detection algorithm SVM-TIA based on SVM and target item analysis method is proposed. The method uses an adaptive method to fitrtificial synthetic samples for SVM boundary samples to alleviate class imbalance. Experimental results show that the proposed algorithm iprovesthe detection results of the recall and precision in some area, and assure a better recommendation quality of recommender systems.⑤ According to statistical distribution of shilling attack ratings and time stamps of malicious ratings information, a shilling attack detection algorithm TS-TIA is proposed based on the target item analysis and time series. By constructing a model based on time series, maximize the difference between time windows with shilling attack ratings and time windows without shilling attack ratings. By the time the project on scoring sequence modeling, so that the normal window containing the sample mean and sample entropy maximization prop attack scores window. Locate the time windows that contain shilling attack ratings, in the rating matrix consits of users, items and rating of users on items, target item analysis method is used to filter out genuine profiles and detect the shilling attack ratings. Experiments show that the performance of TS-TIA based on time series analysis and target item analysis method is effective when shilling attack ratings are concentrated, with relatively low time-consuming.