推荐系统中托攻击防御方法研究

发布时间：2019-02-28 08:49

【摘要】：伴随互联网的普及与电子商务的快速发展,信息数据量以指数级别增长的同时带来了“信息过载”问题。推荐算法通过数据挖掘、机器学习等方式挖掘海量信息中能够帮助电子商务网站为其客户提供符合个性化需求的决策支撑和信息服务,一定程度上有效的缓解了海量数据问题。但系统自身的公开性、推荐算法本身存在的设计缺陷以及用户的介入性导致系统容易遭受恶意干扰、蓄意攻击等操纵行为。因此,安全性成为推荐系统的关键问题。通常将有目的去伪造、更改评分数据的恶意操作称为用户概貌注入攻击或者托攻击。传统的协同过滤技术已然无法满足推荐系统对高安全性、防御性、准确性等推荐可靠性要求。部分商家向推荐系统中恶意注入攻击用户概貌,对推荐系统结果进行人为干预企图谋取私利。这些恶意操作行为严重危害了推荐系统的安全性。如何检测出托攻击并采取有效的方法来防御托攻击刻不容缓,已成为该领域专家学者重要研究问题。相似度度量是协同过滤算法的核心模块,但易于遭受推荐攻击问题。近年来,信誉模型被融合到推荐流程中,加强协同过滤算法的鲁棒性和推荐精确性。基于目前研究趋势,本文提出了两种改进方法提高推荐系统的防御能力。本文主要创新改进内容如下： (1)基于信息熵相似度的托攻击防御方法在协同过滤相关理论的基础上,针对当前相似度度量方法仅考虑评分矩阵数据的局限性,本文提出信息熵来度量正常用户与恶意用户间评分变化幅度差异。融合信息熵模型作为度量相似度的影响因子,弥补了系统遭受攻击时仅依靠传统相似度不足以区分恶意用户的缺陷性。在皮尔森相关系数基础上,本文提出一种改进的相似度度量方法(E-CF),结合评分变化幅度差异降低注入用户概貌的相似性。实验结果表明,E-CF客观地反映托攻击情况下系统防御性增强,并提高了算法精确性。 (2)融合信任更新机制的防攻击推荐算法研究随着社交网络研究的飞速发展,信任关系网络被广泛应用到个性化推荐算法研究中。考虑到推荐用户在过去的推荐历史中所起到的作用也是一个重要的推荐依据因素,即推荐用户的信任度,引入信任更新机制。通过融合信任度和相似度,建立复合推荐权重模型(TE-CF),以真实评分反馈为手段动态更新复合权重,降低攻击用户概貌对推荐结果的影响。实验结果表明,TE-CF客观地反映攻击情况下系统防御性增强,并提高了算法精确性。基于目前研究现状,本文提出两种托攻击防御解决方案。最后,与现有算法进行实验验证和对比分析,并提出进一步研究内容。
[Abstract]:With the popularization of Internet and the rapid development of E-commerce, the data amount of information increases exponentially and brings about the problem of "information overload" at the same time. By means of data mining and machine learning, the recommendation algorithm can help e-commerce website to provide decision support and information service to its customers, which can alleviate the problem of mass data to a certain extent. However, the openness of the system itself, the design defects of the recommended algorithm itself and the user's involvement result in the system being vulnerable to malicious interference, deliberate attack and other manipulation behaviors. Therefore, security has become the key issue of recommendation system. The malicious action to change the score data is called user profile injection attack or proxy attack. The traditional collaborative filtering technology has been unable to meet the recommendation system for high security, defensibility, accuracy and other recommended reliability requirements. Some merchants inject malicious user profile into the recommendation system and interfere with the result of the recommendation system in an attempt to gain self-interest. These malicious operations seriously endanger the security of the recommendation system. How to detect the support attack and take effective methods to defend the support attack is urgent, which has become an important research issue of experts and scholars in this field. Similarity measurement is the core module of collaborative filtering algorithm, but it is vulnerable to recommendation attack. In recent years, reputation models have been incorporated into the recommendation process to enhance the robustness and accuracy of collaborative filtering algorithms. Based on the current research trend, two improved methods are proposed to improve the defense capability of the recommendation system. The main innovation and improvement contents of this paper are as follows: (1) based on the theory of collaborative filtering, the supporting attack defense method based on information entropy similarity is based on the theory of collaborative filtering, aiming at the limitation of the current similarity measurement method only considering the data of scoring matrix. In this paper, information entropy is proposed to measure the variation of scores between normal users and malicious users. The fusion information entropy model as an influential factor to measure similarity makes up for the defects of malicious users when the system is attacked by traditional similarity. Based on Pearson's correlation coefficient, an improved similarity measure method (E-CF) is proposed in this paper, which reduces the similarity of injected user profile combined with the variation of scoring range. The experimental results show that E-CF objectively reflects the system defensibility and improves the accuracy of the algorithm under the condition of support attack. (2) Research on Anti-attack recommendation algorithm based on Trust Renewal Mechanism; with the rapid development of social network research, trust relation network has been widely used in personalized recommendation algorithm research. Considering the role of the recommended user in the past recommendation history is also an important recommendation basis, that is, the degree of trust of the recommended user, the introduction of trust update mechanism. A composite recommendation weight model (TE-CF) is established by combining trust degree and similarity degree. Real score feedback is used as a means to dynamically update composite weights to reduce the impact of attacking user profile on recommendation results. The experimental results show that TE-CF objectively reflects the system defensiveness and improves the accuracy of the algorithm. Based on the current research situation, this paper proposes two solutions for supporting attack defense. Finally, experimental verification and comparative analysis with the existing algorithms are carried out, and further research contents are put forward.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.3;TP393.08

【引证文献】