推荐系统攻击检测算法的研究

发布时间：2018-01-26 09:49

本文关键词： 协同过滤攻击检测 AP聚类用户概貌概貌特征属性　出处：《电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：电子商务的迅速发展给人们的生活提供了更加丰富的选择,但也使得服务信息呈现“超载”趋势,推荐系统是过滤信息的重要手段,是解决信息超载卓有成效的方法。然而由于系统本身对用户的开放性及灵敏性,使其很容易遭到外界的攻击。部分恶意商家在商业利益的驱动下,刻意地向系统中植入一些伪造的用户概貌来影响推荐系统的准确性。如何对外界攻击进行防御和检测,确保电子商务推荐系统的安全成为近年来信息推荐领域的一个新的研究热点。本文综合分析了国内外有关推荐系统安全性的研究现状,并针对基于协同过滤的攻击检测算法进行了深入研究,主要研究工作如下:1.深入分析了协同过滤算法的基本思想和工作流程;研究推荐攻击的相关问题,理解推荐攻击的策略;根据攻击用户概貌的评分策略对攻击模型进行了分类。将现有经典的攻击检测算法进行了分类,通过实验根据几种标准的攻击模型生成对应的攻击用户概貌植入至原始系统,分析比较了攻击前后不同攻击比例和填充比例对推荐系统平均预测偏离度和命中率的影响情况。2.理解研究基于Hv-score值的UnRAP无监督攻击检测算法,分析算法的基本思想和实现流程。在UnRAP检测算法的基础上,事先对系统中的所有用户进行聚类,并将类中的用户评分进行压缩。针对群体用户而不是单个用户来对UnRAP算法进行改进,得到一种基于UnRAP的群组攻击检测算法AP-UnRAP。改进后的算法充分考虑了攻击用户内部之间的高相似性,寻找目标项目时相对单个用户概貌更加准确。3.结合用户概貌特征属性,提出一种基于AP聚类的混合无监督攻击检测算法AP-Mix。通过将用户原始评分矩阵采用PCA降维,并将主分量信息和用户概貌特征属性进行维度组合,用来表示每个用户的整体评分行为;接着,利用一种自适应AP聚类算法对系统中的所有用户进行群组划分;最后,计算每个群组的平均评分偏离度(GRDMA)来找到攻击用户所在的某个群组,进而检测出植入的攻击用户。AP-Mix用组合后的信息代表用户的完整行为,加大了攻击用户和正常用户的区分度,用户群体划分的效果更好,检测性能越强;且事先不需要知道任何攻击的知识,真正做到了无监督检测。最后,通过实验与现有经典检测算法进行对比来验证本文提出新算法的检测高效性。
[Abstract]:The rapid development of electronic commerce provides more choices for people's life, but also makes service information "overload" trend, recommendation system is an important means of filtering information. It is a very effective way to solve the problem of information overload. However, because of the openness and sensitivity of the system to users, it is easy to be attacked by the outside world. Some malicious businesses are driven by commercial interests. Deliberately implant some fake user profiles into the system to affect the accuracy of the recommendation system. How to defend against and detect external attacks. To ensure the security of E-commerce recommendation system has become a new research hotspot in the field of information recommendation in recent years. And the attack detection algorithm based on collaborative filtering is deeply studied. The main research work is as follows: 1. The basic idea and workflow of collaborative filtering algorithm are deeply analyzed. To study the related problems of recommendation attack and understand the strategy of recommendation attack; The attack models are classified according to the scoring strategy of the attack user profile, and the existing classic attack detection algorithms are classified. According to several standard attack models, the corresponding attack user profile is generated by experiments and implanted into the original system. This paper analyzes and compares the influence of different attack ratio and filling ratio before and after attack on the average predictive deviation and hit rate of recommendation system. 2. Understand and study UnRAP unsupervised attack based on Hv-score value. Detection algorithm. The basic idea and implementation flow of the algorithm are analyzed. Based on the UnRAP detection algorithm, all users in the system are clustered in advance. The UnRAP algorithm is improved by compressing the user score in the class and aiming at the group users rather than the individual users. An AP-UnRAP-based group attack detection algorithm based on UnRAP is proposed. The improved algorithm takes into account the high similarity among the users. When looking for the target item, it is more accurate than a single user. 3. Combine the feature attribute of user profile. An AP-Mix-based hybrid unsupervised attack detection algorithm based on AP clustering is proposed. The user's original score matrix is reduced by PCA. The principal component information and the feature attribute of user profile are combined to represent the overall rating behavior of each user. Then, an adaptive AP clustering algorithm is used to group all users in the system. Finally, the average score deviation of each group is calculated to find the group in which the user is attacked. Furthermore, the embedded attack user. AP-Mix uses the combined information to represent the complete behavior of the user, which increases the degree of discrimination between the attacking user and the normal user, and the effect of user group division is better. The stronger the detection performance is; And we do not need to know any knowledge of attack in advance to achieve unsupervised detection. Finally, the effectiveness of the new algorithm is verified by comparing the experimental results with the existing classical detection algorithms.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.3;TP393.08

【参考文献】