面向关联属性的差分隐私信息熵度量方法

Information Entropy Metric Methods of Association Attributes for Differential Privacy

针对差分隐私非交互式多属性关联的合成数据集发布问题,基于信息熵、汉明失真提出了发布数据集隐私度、数据效用、隐私泄露风险的量化方法.首先,利用互信息量分析属性相关度,并以关联依赖图模型表达属性关联.其次,基于图中关键隐私泄露路径构建马尔可夫隐私泄露链,并结合信息熵提出一种关联属性隐私度量模型及方法,可以有效的度量由关联属性引起的隐私泄露量.最后,通过具体实例验证了模型与方法的有效性,并对比分析了该方法的优势.

Privacy leakage and utility measurement are widely concerned issues in multi-attribute datasets by non-interactive differential privacy publishing.In this paper,we have proposed several quantification methods by using information entropy and hamming distortion to quantify the privacy of published dataset,utility of dataset and risk of privacy leakage.First,we have tailored the existing mutual information concept to analyze the relationship among associated attributes and constructed an associated dependency graph model to analyze their correlations among multi-attribute.After that,we have developed a privacy quantification method based on information entropy and privacy leakage Markov chain,which is generated based on the graph of privacy leakage path that has a valid efficiency measurement of the privacy leakage leading by associated attributes.Finally,to justify the efficiency of the proposed model,we have included an illustrative example and demonstrated the advantage of our method by comparing with other methods.