On t-Closeness with KL-Divergence and Semantic Privacy

Abstract

In this paper, we study how to sanitize the publishing data with sensitive attribute to achieve t-closeness and δ-disclosure privacy under Incognito framework. t-closeness is a privacy measure proposed to account for skewness attack and similarity attack, which are limitations of l-diversity. Under the t-closeness model, the distance between the privacy attribute distribution and the global one should be under the threshold t. Whereas semantic privacy (δ-disclosure privacy) is used to measure the incremental information gain from the anonymized tables. We use the Kullback-Leibler divergence to measure the distance between distributions and discuss the properties of the semantic privacy. We also study the relationship between t-closeness with KL-divergence and semantic privacy, and show that t-closeness with KL-divergence and δ-disclosure privacy satisfy the generalization property and the subset property, which entail us to use the Incognito algorithm. Experiments demonstrate the efficiency and effectiveness of our approaches.