面向社会化推荐的托攻击及检测研究

发布时间：2019-03-26 09:21

【摘要】：随着电子商务零售业的迅猛发展和社交网络营销的兴起,以用户间社交关系作为额外输入的社会化推荐系统成为新的研究方向。社会化推荐系统基于社交关系体现用户间相似性这一假设,对解决传统推荐系统中存在的冷启动问题及提高推荐结果的准确性具有重要作用。但社会化推荐系统天然开放性的特点,使其容易受到托攻击者注入虚假欺骗信息(虚假评分或虚假关系等)的影响。此类攻击称为“托攻击”,托攻击严重影响了推荐结果的公正性和真实性,降低了用户对系统的信任度。社会化推荐系统可以看成是传统推荐系统与在线社交网络结合的产物。现有研究大多关注评分驱动的推荐系统或关系驱动的社交网络中托攻击的检测问题,而较少关注同时受评分和关系驱动的社会化推荐系统可能受到的攻击形式与检测手段。针对现有研究的不足,本文首先对社会化推荐系统中的托攻击者的行为方式进行建模,然后提出用于检测推荐系统与社交网络中虚假欺骗信息的特征提取方法,进而得到社会化推荐系统中的托攻击检测技术。本文分别从以下几个方面展开研究:(1)构建面向社会化推荐系统的托攻击模型,并从攻击成本与攻击效果角度对所提模型进行分析。托攻击模型是托攻击者向系统注入虚假用户概貌的手段。通过分析现有社会化推荐技术的工作原理,归纳出托攻击者可能的攻击形式,从而提出托攻击模型。然后分析攻击模型对推荐结果的影响得到所提托攻击模型对社会化推荐系统的攻击效果。(2)针对评分驱动的推荐系统中的托攻击问题,提出一种基于流行度分类特征的托攻击检测方法。推荐系统中托攻击者通过注入虚假评分影响推荐结果,传统方法大多从托攻击者的评分方式入手,此类方法难以对新形式攻击进行检测。为了解决这个问题,从托攻击者与正常用户不同的项目选择行为入手,分析用户概貌中项目流行度分布存在的差异,得到用于检测推荐系统托攻击的特征提取方法,最后结合分类器对推荐系统中的托攻击进行检测。(3)针对关系驱动的社交网络中的托攻击问题,提出一种基于拉普拉斯得分的托攻击检测方法。社交网络中托攻击者通过注入虚假关系提升自己的影响力,从而达到传播虚假信息的目的。现有方法在训练模型时使用的特征维度较高,造成检测准确性不足。为了解决这个问题,提出无监督的特征选择方法,该方法通过拉普拉斯得分衡量特征的局部信息保持能力,以进行特征选择。在此基础上,结合半监督学习方法对社交网络中的托攻击进行检测。(4)面向社会化推荐系统中的托攻击检测问题,提出一种基于半监督协同训练的社会化推荐系统托攻击检测方法。社会化推荐系统中的用户包括评分特征与关系特征,因此可以利用推荐系统与社交网络中检测托攻击的特征提取方法,得到用户评分视图与关系视图的特征。同时考虑到系统中标签不足问题,将半监督协同训练算法用于模型构建,在两个独立的特征子图上分别训练分类器,从而对社会化推荐系统中的托攻击进行检测。
[Abstract]:With the rapid development of e-commerce and the rise of social network marketing, the social relationship between users as an additional input becomes the new research direction. The social recommendation system is based on the assumption that the social relation reflects the inter-user similarity, and plays an important role in solving the cold start problem existing in the traditional recommendation system and improving the accuracy of the recommendation result. But the nature of the self-opening of the social recommendation system makes it vulnerable to the influence of the false-lying information (false score or false relationship) injected by the attacker. This kind of attack is called "to attack", and the support attack seriously affects the fairness and the authenticity of the recommendation result, and reduces the user's trust in the system. The social recommendation system can be regarded as a product of the traditional recommendation system and the online social network. The existing research focuses on the detection of the support attack in the recommendation system or relationship driven by the score drive, while the less attention is paid to the form of attack and the means of detection of the social recommendation system driven by the scoring and the relationship. In view of the shortcomings of the existing research, this paper firstly models the behavior of the attacker in the social recommendation system, and then puts forward a feature extraction method for detecting the false spoofed information in the recommendation system and the social network, And then the carrier attack detection technology in the social recommendation system is obtained. This paper studies from the following aspects: (1) to construct a support attack model for the social recommendation system, and to analyze the proposed model from the attack cost and the attack effect angle. The support attack model is a means to allow an attacker to inject a false user profile into the system. Based on the analysis of the working principle of the existing social recommendation technology, the possible attack form of the attacker is summarized, and the attack model is put forward. Then the influence of the analysis attack model on the recommendation result is obtained, and the attack effect of the proposed attack model on the socialization recommendation system is obtained. (2) In order to solve the problem of support attack in the recommendation system driven by the score, a method for detecting the support attack based on the feature of the popularity classification is proposed. In the proposed system, it is difficult to detect the new forms of attack by injecting false scores to influence the recommended results. in ord to solve that problem, starting with different project selection behaviors of an attacker and a normal user, the difference in the distribution of the project popularity in the profile of a user is analyzed to obtain a feature extraction method for detecting a recommended system to attack, And finally, combining the classifier to detect the support attack in the recommendation system. (3) To solve the problem of the support attack in the social network driven by the relation, a method for detecting the support attack based on the Laplacian score is proposed. An attacker in a social network raises his influence by injecting a false relationship, thereby achieving the purpose of propagating false information. The method has the advantages that the characteristic dimension used in the training model is high, and the detection accuracy is insufficient. In order to solve this problem, a non-supervised feature selection method is proposed, which measures the local information retention capability of the feature by the Laplacian score for feature selection. On this basis, a semi-supervised learning method is used to detect the support attack in the social network. (4) In order to solve the problem of support attack detection in the social recommendation system, a method for detecting the support attack of the socialization recommendation system based on the semi-supervised cooperative training is proposed. The user of the social recommendation system includes the score feature and the relationship feature, so that the feature of the user's scoring view and the relationship view can be obtained by using the recommendation system and the feature extraction method of detecting the tray attack in the social network. At the same time, the semi-supervised cooperative training algorithm is used for the model construction, and the classifier is trained on two independent feature subgraphs, so that the support attack in the social recommendation system is detected.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【参考文献】