基于OLH的效用优化本地差分隐私机制

Utility-optimized Local Differential Privacy Mechanisms Based on OLH

效用优化本地差分隐私模型能够在保证隐私的前提下提高估计结果准确度.但现有的效用优化本地差分隐私协议存在着数据效用低或通信代价大的问题.本文针对现有效用优化本地差分隐私协议难以兼顾低通信代价和高数据效用的不足,基于OLH(optimized local hashing)协议提出了符合效用优化本地差分隐私模型的uOLH(utility-optimized OLH)协议.该协议在原始数据定义域很大时,同时具有低通信代价和高数据效用的特点,兼顾u RR(utility-optimized randomized response)和uRAP(utilityoptimized randomized aggregatable privacy-preserving ordinal response)二者优势.本文进一步考虑了用户的个性化隐私保护需求,构造了优化加权组合机制DWC(data weighted combination),在此基础上提出了个性化效用优化本地差分隐私协议uOLH-DWC.允许用户自由选择隐私级别,并能够提升估计结果的准确度,可输出多个隐私级别下的频率估计结果.在真实和模拟数据集上的实验结果表明,u OLH协议可以同时具有低通信代价与高数据效用,且uOLH-DWC协议可令用户自由选择隐私预算,并提升了各个隐私级别下估计结果的准确度.

ULDP(utility-optimized local differential privacy) model can improve the accuracy of estimation results on the premise of ensuring privacy. However, the existing ULDP protocols suffer from low data utility or high communication cost. To overcome the disadvantages of u RR(utility-optimized randomized response) and uRAP(utility-optimized randomized aggregatable privacy-preserving ordinal response), based on the OLH(optimized local hashing) protocol, this paper proposes a uOLH(utility-optimized OLH) protocol in ULDP model. When the original data has a large domain, the uOLH protocol has a low communication cost similar to uRR and high data utility similar to uRAP.This paper further considers privacy protection of users’ data, and constructs an optimized weighted combination mechanism DWC(data weighted combination). Based on the DWC mechanism, a personalized ULDP protocol u OLH-DWC is designed which allows users to freely choose their privacy budget. The uOLH-DWC can generate frequency estimation results with the smallest error and output frequency estimation results of multiple privacy levels to satisfy data users with different levels of credibility. Experimental results on real and simulated data sets show that the proposed uOLH achieves both low communication cost and high data utility. The proposed u OLH-DWC allows users to freely choose their privacy budget and improves the data utility of the estimated results at each privacy level.