面向敏感数据逐步发布的差分隐私可调高斯机制研究

发布时间：2018-10-09 10:57

【摘要】：差分隐私(Differential Privacy,DP)是一种新型的隐私保护技术,它通过对原始查询结果添加噪音后发布出去来达到隐私保护的目的。差分隐私下的隐私保护水平由隐私保护预算ε决定。针对以往的差分隐私数据发布后无法通过调整隐私保护预算以满足用户动态对隐私保护水平变化的需求,论文提出了一种面向敏感数据逐步发布的差分隐私可调高斯机制。其主要研究内容包括:(1)提出可调高斯机制。证明了在统计数据的发布过程中,对原始数据添加服从高斯分布的噪音1V,发布满足1 1(ε,δ)-差分隐私的结果1y,若将隐私保护预算1ε调整到2ε2 1(εε),在噪音1V的基础上添加噪音2V,发布更精确的数据结果2y,则最终的发布结果1 2(y,y)满足2 2(ε,δ)-差分隐私。(2)给出了可调高斯机制下数据发布的过程。采用可调高斯机制将隐私保护预算从1ε放大到2ε,并将隐私保护预算的一次放大扩展到多次放大,在数据发布的过程中,若对原始统计数据添加具有马尔可夫性质的高斯噪音,即下一次添加多少噪音只与当前添加噪音的多少有关,与过去添加的噪音量无关,则可以实现隐私保护预算的多次放大,从而达到差分隐私下敏感数据的逐步发布。(3)可调高斯机制的评估。基于差分隐私下隐私损失的定义,比较了可调拉普拉斯机制、高斯组合机制、可调高斯机制等机制的隐私损失,发现采用可调高斯机制调整隐私保护预算的隐私损失小,验证了采用可调高斯机制放大隐私保护预算具有一定的优越性。论文对可调高斯机制下的数据发布进行了实验分析,采用该机制放大隐私保护预算的过程中,发布数据的均方误差越来越小,即用户获得的数据结果精度越来越高。该研究面向一般意义的(ε,δ)-差分隐私,根据用户的隐私需求调整隐私保护预算来发布数据的,使得隐私数据保护和使用之间更易于平衡,有利于推动差分隐私的应用,促进大数据技术的发展。
[Abstract]:Differential privacy (Differential Privacy,DP) is a new privacy protection technology. It can achieve the purpose of privacy protection by adding noise to the original query results and publishing them out. The level of privacy protection under differential privacy is determined by the privacy protection budget 蔚. In view of the past differential privacy data can not be adjusted by adjusting the privacy protection budget to meet the needs of users to change the level of privacy protection, this paper proposes a differential privacy adjustable Gao Si mechanism for the gradual release of sensitive data. The main research contents are as follows: (1) the adjustable Gao Si mechanism is proposed. Proves that in the process of releasing statistical data, To add noise from Gao Si distribution to original data 1V, release the result that satisfies 11 (蔚, 未) -difference privacy 1y. if we adjust the privacy protection budget 1 蔚 to 2 蔚 21 (蔚), add noise 2V on the basis of noise 1V, release more accurate data result 2y. then The final result 12 (YY) satisfies 22 (蔚, 未) -difference privacy. (2) the process of data release under the adjustable Gao Si mechanism is given. The adjustable Gao Si mechanism is used to enlarge the privacy protection budget from 1 蔚 to 2 蔚, and the one-time enlargement of the privacy protection budget is extended to multiple magnification. That is, how much noise is added next time is only related to how much noise is added at present, and not related to the amount of noise added in the past, so that the privacy protection budget can be magnified several times. Thus, the sensitive data can be released step by step under differential privacy. (3) the evaluation of adjustable Gao Si mechanism. Based on the definition of privacy loss under differential privacy, the privacy loss of adjustable Laplace mechanism, Gao Si combination mechanism and adjustable Gao Si mechanism are compared. Verify that the adjustable Gao Si mechanism to enlarge the privacy protection budget has some advantages. This paper makes an experimental analysis on the data release under the adjustable Gao Si mechanism. In the process of enlarging the privacy protection budget by this mechanism, the mean square error of the published data becomes smaller and smaller, that is, the accuracy of the data obtained by the user is getting higher and higher. This research aims at (蔚, 未) -differential privacy in general sense, adjusts the privacy protection budget to release data according to the user's privacy needs, makes the balance between privacy data protection and use easier, and promotes the application of differential privacy. To promote the development of big data's technology.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP309

【参考文献】