A Research Of Shilling Attack Detection Based On Feature Selection

As a kind of important personalized service system, collaborative filtering recommendation system is more and more widely used in the field of e-commerce and other internet domains, provides convenience for users. However, with the emergence of shilling attack, the security of recommendation system is seriously threatened, which restricts the development of electronic commerce system to some extent. In order to resist the destruction of shilling attack, research on its detection has received wide attention. Based on the research status at home and abroad, this paper discusses the detection method of shilling attack further.Because the shilling attack detection has timeliness, the subsequent detection are needed. The methods for initial detection and the subsequent detection should be different. The initial detection needs to be as accurate as possible, while the subsequent detection needs algorithm with high efficiency. Therefore, this paper researches the shilling attack detection algorithm according to the above two kinds of circumstances respectively. Under this premise, this paper divides the design problem of detection algorithm into two sub-problems: How to select valid feature indexes and how to design suitable detection algorithm on this set of feature indexes.Aim at the problems,on considering that traditional detection algorithm of shilling attack can not deal with the various types of shilling attacks, feature selection algorithm of detection index based on information entropy dynamics selection for the current data set is first put forward. Combined with characteristics of information entropy, divisions of normal users and attack users are regarded as two kinds of random events and information entropy of all kinds of characteristic indexes is figured out. Trough this, detection ability of characteristic indexes is valued and the appropriate characteristic index is chosen out.Then, unsupervised detection algorithm based on the degree of outlier is designed, which identifies attack users from the outlier angles of users' profile features. Different models of shilling attacks are respectively constructed on Movie Lens data set and detected by application algorithm to verify the validity of the algorithm. Then it is compared with other several mainstream detection algorithms to complete the verification on the accuracy of the method.After that, according to high efficiency demand for subsequent detection, a set of shilling attack feature extraction methods based on project real popularity are raised on the front basis of getting the real user profile sets. Because the definition of popularity is based on statistics item scores, the computation complexity will be greatly reduced. Because the definition of popularity is based on statistics item scores, the computation complexity will be greatly reduced. Taking the selection differences of scoring items between real users and attack users as the breakthrough point, two characteristic indexes of users' average project popularity and its information entropy are summed up. Combined these two characteristic indexes with previous detection algorithm, the subsequent duplicate detection of shilling attack is accomplished.Finally, a simulation system of shilling attack is actualized based on the previous algorithm design. By imitating the whole process of actual shilling attack detection, the validity of the set of detection methods in this passage is proved.