Research On The Defense Of Shilling Attack In Recommendation System

With the rapid development of the Internet and E-commerce, the information increases exponentially, and brings along with the problem of "information overload". By using the data mining, machine learning and so on methods to exploit the massive information on the e-commerce sites, recommendation algorithm can offer personalized decision support and information services to its customers, and effectively alleviate the information overload problem. But due to the openness of the system, the participation of the users and the inherent design deficiencies of recommendation algorithm make the system vulnerable to manipulative behaviors like malicious interference and attack. Thus, security is a key issue of the recommendation system. Generally, the manipulative behaviors that faking and modifying the rating data are called the "user’s profile injection attack" or the "shilling attack". The traditional collaborative filtering algorithms cannot meet the requirement of a reliable recommendation system in terms of high safety, defensiveness and accuracy. Some businessman exerts malicious injection attacks on the recommendation system and interfere the recommendation results. These malicious acts seriously endanger the safety of the recommendation system. How to detect the shilling attacks and take immediately effective methods to defence these attacks has become an important research problem.Similarity metrics is the core module of the collaborative filtering algorithms. However, it is vulnerable to the recommendation attacks. Recently, the reputation model is intergrated into the recommendation attacks to enhance the robustness and accuracy of the collaborative filtering algorithms. Basing on this research trend, we propose two improved methods to protect the recommendation system from shilling attack. The details are as follows:(1)Information entropy similarity based method.Basing on the relevant theories of collaborative filtering and limitations of the current similarity measure based recommendation methods, we provide an information entropy based metric to measure degree of changes in user comments, by analyzing the differences between normal users and malicious users. The integrated information entropy model maks up the limitations of the system, such as it can not distinguish normal users from malicious users in an attackd. Basing on Person correlation coefficient, we propose an improved similarity measure (E-CF in short) to reduce the similarity of injected users profile according to the rating differences. The experimental results show that, E-CF is robust to the shilling attack and improve the performance of the recommendation system.(2) Anti attack recommendation algorithm integrated with trust update mechanismTrust network is a widely used technology in the existing personalized recommendation systems. Based on the observation that the effect exerted on the recommendation history by the recommender users is an important recommendation reference factor, we propose an improved method by combing the trust network and similarith weights, and establish a composite model(TE-CF in short). The experimental results show that, TE-CF is robust to the shilling attack and improve the performance of the recommendation system.Based on the analysis of the existing researches, this paper proposed two works for the defence of Shilling attack problem in recommendation system. The defense methods are experimental verified and compared with several existing algorithms, and some futher work are proposed accordingly.