基于元胞自动机的欺诈团伙检测模型研究

发布时间：2018-03-14 15:38

本文选题：欺诈检测　切入点：元胞自动机　出处：《重庆大学》2014年硕士论文　论文类型：学位论文

【摘要】：C2C平台中存在不诚实的用户，他们通过产生大量的虚拟交易快速提高信用积分，然后在高信用度的掩护下实施欺诈性质的交易，让C2C电子商务遭遇了严重的信用危机。先刷信用积分再实施欺诈的行为通常由团伙产生，团伙由欺诈性卖家和虚假的买家同伙组成，他们相互掩护使得对他们的识别难度非常大。有效地对C2C平台用户可信性进行重新评估并识别欺诈团伙，能维持C2C公平的交易环境，缓解信用危机，也能给消费者减少麻烦和损失，具有极大的研究意义。目前对在线交易欺诈的研究主要集中在商品拍卖、股票和期货市场，对一口价商品交易欺诈的研究非常少。我们仔细分析了国内外欺诈检测相关文献，发现当前使用的检测方法存在着一些问题，在此基础上，本文以发掘识别欺诈团伙新途径为目的，寻求一种既考虑用户基本特征属性又考虑用户所处的局部交易网络的全新的检测模型。元胞自动机（CA）能以微观个体简单的局部自组织行为表现系统整体复杂性，不规则元胞自动机（ICA）是对标准CA的扩展，能对复杂交易网络进行模拟，而学习自动机（LA）能根据环境反馈自动调整自身状态，将它们结合在一起形成了一个具有强大适应能力的能对复杂交易网中用户状态进行判别的分类模型FD_ICLA。本文采用机器学习算法，基于用户基本属性及交易统计属性挖掘产生本地规则。本地规则以邻居相关信息和内嵌LA选择的动作为输入产生加强信号，内嵌LA依据此信号调整元胞状态。FD_ICLA模型采用“自下而上”的模拟方法，通过微观上反复执行的推理，实现对宏观状态的判定。用单个FD_ICLA进程对包含上百万个节点的交易网络进行分析是非常耗时的，考虑到元胞自动机的局部依赖性，本文基于图的K划分算法，提出了并行FD_ICLA模型，，该改进模型能有效地将计算压力分散到多个的机器，增强了模型的扩展能力。同时，本文基于Gephi实现了可视化原型系统能直观展示模型分析结果。最后，为了检验模型对欺诈团伙的识别效果及时间性能，本文从Kongfz平台采集真实交易数据集，并组织多组对比实验。实验结果表明：1）相对S2C+SNA欺诈检测算法及PeerGroup欺诈检测算法，FD_ICLA模型能以更高的精确度对同盟进行识别，而且能更有效的挖掘交易网中存在的欺诈团伙；2）并行FD_ICLA模型能有效弥补单进程模式高耗时缺陷。
[Abstract]:There are dishonest users in the C2C platform, who increase credit score quickly by generating a large number of virtual transactions, and then carry out fraudulent transactions under the cover of high credit degree. C2C e-commerce has suffered a serious credit crisis. Credit points are used before fraud is committed by gangs, which are made up of fraudulent sellers and false buyers. They cover each other and make it very difficult to identify them. Effectively reassessing the credibility of C2C platform users and identifying fraudulent gangs can maintain a fair trading environment for C2C and alleviate the credit crisis. Also can reduce the trouble and loss of consumers, has a great significance of research. At present, the research on online trading fraud is mainly focused on the commodity auction, stock and futures markets, and the research on the fraud of one price commodity trading is very few. We have carefully analyzed the domestic and foreign relevant documents on fraud detection. It is found that there are some problems in the current detection methods. On this basis, the purpose of this paper is to find new ways to identify fraudulent gangs. This paper seeks a new detection model which considers both the basic characteristics of the user and the local transaction network in which the user is located. Cellular automata (CAA) can represent the whole complexity of the system by the simple local self-organization behavior of micro-individuals. The irregular cellular automata (ICA) is an extension of the standard CA and can be used to simulate complex trading networks. Learning automata can automatically adjust their state according to environmental feedback. This paper combines them to form a powerful adaptive classification model FDS _ CLAs, which can judge the user's status in a complex trading network. In this paper, a machine learning algorithm is used. Local rules are generated based on user's basic attributes and transaction statistical attributes. Local rules generate reinforcement signals based on neighbor related information and embedded LA selected actions. Based on this signal, the embedded LA adjusts the cellular state. FDC _ ICLA model adopts the "bottom-up" simulation method, and realizes the judgment of macroscopic state by microcosmic repeated reasoning. It is time consuming to analyze a transaction network with millions of nodes by using a single FD_ICLA process. Considering the local dependence of cellular automata, this paper proposes a parallel FD_ICLA model based on the K partition algorithm of graph. The improved model can effectively disperse the computational pressure to multiple machines and enhance the expansion ability of the model. At the same time, this paper implements a visual prototype system based on Gephi, which can visually display the analysis results of the model. Finally, in order to test the effect and time performance of the model, we collect the real transaction data set from Kongfz platform. The experimental results show that compared with S2C SNA fraud detection algorithm and PeerGroup fraud detection algorithm, the model can identify the alliance with higher accuracy. Moreover, the parallel FD_ICLA model can effectively mine the fraud gang in the transaction network. The parallel FD_ICLA model can effectively compensate for the high time consuming defects of the single process pattern.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F724.6;TP301.1

【参考文献】