Shilling Attack Detection Algorithm Combined With Time Dimensions

With the explosive growth of information resources, personalized recommendation technology has been widely used as a tool to effectively alleviate the information overload. However, due to recommendation systems using collaborative user data to generate recommendation, the attackers can inject fake ratings into the recommendation systems to influence the recommendation lists, to achieve the purpose of attack. The injection of fake ratings will seriously affect the effectiveness of the recommendation systems. Therefore, it is significant to research on the detection of the shilling attack.The existing shilling attack detection methods mainly include the detection of fake users and abnormal items. The fake user detection distinguishes real users and fake users by computing user profile's properties. The method for specific attack model can obtain good detection results, but they are difficult to detect mixed or new attack models. The abnormal item detection is mainly based on finding the abnormal interval of item's rating time series, but it is difficult to detect the fake users.These two kinds of detection methods consider the shilling attack only from the perspective of users or items. In this paper, we combine these two kinds of methods to form a new shilling attack detection method from both user's and item's perspective. The method first obtains a suspicious user set by using the item anomaly detection approach based on time series. Secondly, the method uses the fake user detection approach based on user profile features to obtain another suspicious user set. Finally, the method integrates the two suspicious sets to achieve shilling attack detection.The main contents of this paper are as follows:(1) It analyzes of the research background, significance and status of recommendation system, and shilling attack, and illuminates the research content of this paper. Then this paper introduces the related work of mainstream recommendation algorithms, shilling attack models and shilling attack detection algorithms.(2) The injection time as a critical factor of the shilling attack is less mentioned in the research of attack models. This paper considers both rating and rating time dimensions. According to the different strategies of injection time, this paper proposes the shilling attack models which combine attack time with ratings.(3) This paper analyzes the detection results of the method “item anomaly detection based on dynamic partition for time series'. Then the paper compares various abnormal item detection methods based on time series.(4) This paper detects shilling attack from the perspective both fake users and abnormal items. It takes advantage of the item anomaly detection method can successfully detect the mix attack models and not restrict by the specific attack model. The proposed method(R-RTS) combined the rating deviation from mean agreement(RDMA) and time series to detect fake users. This method combines RDMA with item anomaly detection method based on dynamic partition for time series to improve the detection results and universality. This paper analyzes the detection results of R-RTS in different attack conditions. Then the paper conducts a comparison of R-RTS with two new shilling attack detection algorithms to further evaluate the validity of the method and analyzes the experimental results and the time complexity in detail.