Researches On Push Shilling Attacks In Collaborative Filtering Recommender Systems

The problem of information overload makes collaborative filtering recommendation system developed rapidly, and it is widely used in the field of application. To provide effective personalized information services for the customer terminal is also facing its own security issues. User records as the basis to construct different supporting attack model in limited knowledge forged user profile and quantitative injection database recommendation system prediction score and the user generated false recommendation information acquisition unfair economic interests. Therefore, it is imperative to maintain the security of collaborative filtering recommendation system, and also caused widespread concern in the field of information security, and put forward different support attack detection methods.The detection algorithm of the support attack can be divided into the basis of supervision, unsupervised and semi supervised learning. Based on supervised learning algorithm for new or hybrid attacks, the algorithm is based on the assumption that the unsupervised algorithm is highly similar to the attacker. However, the database system has a precious mark user records, it is the system can trust records, these users record true and credible, there is no false, can truly reflect the degree of user preferences, therefore, this paper proposes a semi supervised detection algorithm based on feature index, which is based on the rule of marking user data distribution to further improve the accuracy of the detection of support attacks. Main work includes:Dependent marker user center design mean and user similarity(MUS). The similar user clustering into a cluster, and then determine the user profile is to prop up the attack. Marker clustering center users to input user clustering, except for the attack input users mean scores and users to mark the mean score of difference are similar, so as to find out the target, for an eventual utilization index characteristics distinguish attack profile provides effective classification center initial value. Clustering process using K-Means algorithm, given the initial cluster center number algorithm, described the initial center selection method, and further improve the classification results. In order to avoid detection, the attacker will use some strategies to build a new model, which makes the attack cost less, and the attack effect is better. Confusion strategy is introduced which leads to the reduction of the difference between the attack profile and normal profile, so as to avoid the shilling attack detection. By injecting noise and pop score offset, the target score offset confusion strategy transform attack profile and popular as attributes for classification of and cross design a new attack model, confused popular cross attack model. Model reduces the difference between the real user profiles, makes it easier to become the user’s neighbor. Designed to confuse popular cross attack profile generator automatically generates attack profile, injection profile attack that the effectiveness of the attack model, show mixed support attack comparing the basic supporting attack hazard model more.From user labeled screening to obtain high degree of trust users tag filtering popular effects, to calculate the high trust user signatures index value matrix, using the intra class scatter matrix of the single pattern feature extraction method of calculation of the high trust degree of user signatures index value of the feature vector. The use of high trust degree mark user feature value of feature vectors to user input characteristic index value vector is compressed and then recognize the attack profile, show that the detection algorithm of mixed support attack model in different filling scale can better detect shilling attacks.