Research On User Profile Injection Attack Detection Algorithm For Recommender Systems

With the rapid development of e-commerce, we come into the era of networkeconomy, but it also brings the problem of “information overload”, recommendationtechnology was proposed forward to solve this problem. However, due to the inherentopenness of the recommendation system and sensitivity to user’s information, making itvery vulnerable to the impact of user profiles injection attacks, this attack seriouslydegrade the quality of the recommended. Therefore, the protection of the safety of therecommended system is a urgent problem. In this paper, on the basis of the research andanalysis about the present situation at home and abroad, has a further deep research onthe users attack profile injection attack detection algorithm in the collaborative filteringrecommendation system.Firstly, aiming at the problem that the precision of the existing user profile injectionattack detection algorithms is not high, by analyzing the differences between thecharacteristics of the attack user profiles and normal user profiles filling degree, based onSVM and rough set theory we propose an approach to detect user profile attack. First ofall, we use SVM to perform the detection based on combing the fill size with the existingattributes and the proposed attributes respectively. The detection results are used togenerate the information table. Then, we use the decision rules generated by rough settheory to determine the information table to generate the final test results.Secondly, existing attack detection algorithms cannot detect the new model ofrecommendation attack effectively. Thus, from the degree of new model recommendationattack constantly appeared in the practical application, an attack detection algorithmbased on cluster of user profiles injection is proposed. The algorithm first extracts somecharacteristics which do not depend on the attack model, then using the k-meansclustering algorithm on user profiles’ characteristics. Also we use random samplingtechnique from the ratings database randomly select the appropriate amount of realprofile as seed data sets, depend on the number and proportion user profile come fromseed data sets in the clusters, we could determine the attack user cluster. In the end, we use the characteristic of user-item average rating deviation degree determined the targetitem, and detect the corresponding attack profiles.Finally, we give the experimental evaluations and analysis of the algorithmsproposed in this paper, compare the performance between the proposed algorithms andother existing algorithms, and make the conclusions and prospects for the further search.