标题：Detecting Malicious Behavior and Collusion for Online Rating System
作者：Cao, Liu; Sun, Yuqing; Wang, Shaoqing; Li, Mingzhu
通讯作者：Sun, YQ;Sun, YQ
作者机构：[Cao, Liu; Sun, Yuqing; Wang, Shaoqing] Shandong Univ, Sch Comp Sci & Technol, Jinan, Peoples R China.; [Li, Mingzhu] Shandong Univ, Sch Software En 更多
会议名称：15th IEEE Int Conf on Trust, Security and Privacy in Comp and Commun / 10th IEEE Int Conf on Big Data Science and Engineering / 14th IEEE Int Symposium on Parallel and Distributed Proc with Applicat (IEEE Trustcom/BigDataSE/ISPA)
会议日期：AUG 23-26, 2016
来源：2016 IEEE TRUSTCOM/BIGDATASE/ISPA
摘要：Since the collective ratings and reviews on an online rating system have a high influence on user decisions, there exist more and more fraudulent behaviors, such as "ballot stuffing" and "bad-mouthing". Most of the current methods solve these problems by detecting one or a few malicious patterns. But they cannot solve these patterns simultaneously and cannot solve the case where dominating population on an item is malicious. In this paper, we investigate malicious behaviors and collusion from three aspects: the normal standards are learned from crowd behaviors rather than predefined patterns; a behavior anomaly is considered from both current and historical views; and the negative influence is also taken into account. We propose a set of metrics to comprehensively detect malicious behavior patterns. The User-Deviation metric detects how much a user is deviated from the crowds. The Behavior-Turbulence metric detects how different a user is from his/her historical behaviors, and the User Behavior Influence metric measures how much negative influence a user may cause. By orchestrating these statistical metrics, different user anomaly behaviors can be recognized. Compared with previous methods, our metrics can not only solve the problem that the major population is malicious, but also avoid misjudging characteristic people. Besides, we propose the collective behavior influence metric. and coherence metric so as to evaluate collusion of a group of people. Our method is verified against some real datasets and the results show that it outperforms the existing models.