【作者】 王宇；

【导师】 陈志立；

【作者基本信息】 安徽大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 随着大数据时代的来临,网络中信息过载的问题日益突出,用户往往很难从海量的数据中快速定位自己需要的资源。推荐算法通过挖掘用户和信息产品之间的二元关系很好的解决了这一问题,但是由于推荐算法在给用户进行推荐的过程中,往往需要使用用户的历史信息,这对用户的隐私产生了巨大威胁。差分隐私作为近几年保护隐私的一种新的方式,以高隐私保护效率,安全性可证明,安全级别可控等优点受到广泛关注。目前关于差分隐私的研究主要集中在数据发布方向,但是差分隐私的优点使得它同样十分适合用来解决推荐系统中的用户隐私保护问题。因此,在推荐系统中应用差分隐私来保护用户隐私的研究是十分有意义的。本文通过调研现有的差分隐私保护技术和经典的推荐算法,着重探究了应用差分隐私保护技术来解决协同过滤算法中的隐私问题,主要贡献包括:(1)针对协同过滤推荐算法中的用户隐私保护问题,本文基于差分隐私保护技术提出了一种高效可行的隐私保护方法(KDPCF)。具体做法是先利用K-means聚类算法将所有用户进行聚类,然后对聚类的结果进行优化,最后在优化后的结果上进行基于用户的协同过滤推荐。为了保护用户的隐私,在推荐过程中我们引入了差分隐私的指数机制。同时为了降低差分隐私对推荐准确率的影响,本文中指数机制直接输出整个邻居集合而不是多次使用指数机制逐个输出邻居,这样做大大降低了指数机制的使用次数,避免方案中引入大量不必要的噪声,提高了推荐准确性。最后,我们在理论上证明我们的算法满足差分隐私的安全性的前提下,使用真实数据进行了对比实验,实验结果证明我们的方案是可行的。(2)本文设计了一个基于推荐代理的用户隐私保护方案(RLAS)。在本地差分隐私保护的基础上,结合不同的隐私保护方法的优点,使得在保证用户隐私安全的同时兼顾推荐系统的准确性。具体来说,本文首先从同一批请求推荐服务的用户群中选出一个用户作为代理,其他用户将自己匿名化的用户数据发送给此代理。代理需要使用本地差分隐私扰动自己的隐私,插入到匿名用户数据集中,最后统一发送给推荐服务方。在整个推荐过程中用户不直接和推荐服务方进行交互,而是让代理与推荐服务方进行沟通,这样做的好处是降低了用户隐私泄露的风险。同时我们为了保护代理用户的隐私,需要使用本地差分隐私来进行扰动处理。最后我们仍然通过实验来验证我们的想法,实验结果也确实达到了我们的预期。更多还原

【Abstract】 With the advent of the era of big data,the problem of information overload has become increasingly prominent in the network,and it is often difficult for users to quickly locate the resources they need from massive amounts of data.The recommendation algorithm solves this problem by mining the binary relationship between the user and the information product.However,the recommendation algorithm often needs to use the user’s historical information in the process of recommending to the user,which poses a great threat to the user’s privacy.As a new way to protect privacy in recent years,differential privacy has attracted wide attention due to its high privacy protection efficiency,provable security,and controllable security level.The current research on differential privacy focuses on the direction of data distribution,but the advantages of differential privacy make it equally suitable for solving user privacy protection problems in recommendation systems.Therefore,it is meaningful to apply differential privacy protection to protect user privacy in recommendation system.By investigating the existing differential privacy protection technology and the classic recommendation algorithm,this paper focuses on the application of differential privacy protection technology to solve the privacy problem in the collaborative filtering recommendation algorithm.The main contributions include:(1)Aiming at the problem of user privacy protection in collaborative filtering recommendation algorithm,this paper proposes an efficient and feasible privacy protection method based on differential privacy protection technology.The specific method is to use K-means clustering algorithm to cluster all users,then optimize the clustering results,and finally perform user-based collaborative filtering recommendation on the optimized results.In order to protect the privacy of users,we introduced the exponential mechanism of differential privacy in the recommendation process.At the same time,in order to reduce the impact of differential privacy on the recommendation accuracy,the exponential mechanism directly outputs the entire neighbor set instead of outputting neighbors one by one.This practice greatly reduces the number of times the exponential mechanism is used,avoids introducing a large amount of unnecessary noise into the scheme,and improves the accuracy of the recommendation.Finally,we theoretically prove that our algorithm satisfies the security of differential privacy,and uses real data to conduct comparative experiments,and the experimental results prove that our scheme is feasible.(2)This paper designs a user privacy protection scheme(RLAS)based on the recommendation agent.On the basis of local differential privacy protection,combined with the advantages of different privacy protection methods,the accuracy of the recommendation system can be taken into account while ensuring the privacy security of users.Specifically,this article first selects a user from the same group of users requesting recommendation services as a agent,and other users send their anonymized data to it,the agent needs to use local differential privacy to perturb own privacy,insert it into the anonymous user data set,and finally send all data set to the recommendation service provider.During the entire recommendation process,the user does not directly interact with the recommendation service provider,but allows the agent to communicate with the recommendation service provider.This has the advantage of reducing the risk of user privacy leakage.At the same time,in order to protect the privacy of the agent user,we use local differential privacy to perform the disturbance processing.Finally,we still experimented to verify our ideas,and the experimental results did meet our expectations.更多还原