差分可辨性隐私参数的迭代分配方法

Iterative Allocation Method of Privacy Parameter for Differential Identifiability

在大数据时代,对个人隐私的保护不容忽视.于2013年被提出的ρ-差分可辨性定义解决了传统差分隐私仅关注个体对数据库输出影响的问题,使隐私保护的重点转移到防止个体被重新识别上,更加符合相关法律的定义.然而,现阶段对差分可辨性的相关研究较少.本文基于差分可辨性组合性质,提出了差分可辨性隐私参数的迭代分配方法,能够在迭代轮数固定和未知两种情况下分配差分可辨性隐私保护参数,使最终模型满足差分可辨性的隐私定义.对于某些需要迭代的模型,如聚类算法k-means,在聚类过程中可能出现隐私泄露,可以借助差分可辨性的实现机制来对每轮迭代进行加噪处理来保护隐私.实验结果表明,本文方法对数据进行噪声添加后,一定程度上能够保证经过差分可辨性隐私保护的聚类结果可用性.

The privacy protection in big data era is an important problem. The definition of ρ-differential identifiability proposed in 2013 solves the problem that traditional differential privacy does not focus on the attacker’s background knowledge but only focuses on the individual’s impact on the database output, which shifts the focus of privacy protection to the protection of re-identifying individuals, and it meets the legal requirements of privacy. So far the research on differential identifiability is limited. This paper introduces an iterative privacy parameter allocation method of differential identifiability based on its combination properties. Each round of iteration is given privacy parameter allocation method, and can be used in case of fixed number of iteration rounds or unknown number of rounds. For some iterative model, such as k-means clustering algorithm, privacy may be disclosed in the process of clustering. With the help of the implementation of differential identifiability, each round of iteration can add noise to protect the privacy. The final model satisfies the definition of differential identifiability. Experimental results show that the designed scheme with noise data addition can guarantee the availability of clustering results.