V-dispersed synthetic data based on a mixture model with constraints

Abstract

In this paper a new approach is proposed for the generation of synthetic microdata which reduces attribute disclosure for continuous variables. First, we define a metric of attribute disclosure which is called v-dispersion. This metric quantifies the risk based on the volume of the multidimensional confidence regions for the original data values. Next we describe a method that satisfies the requirements of v-dispersion. This method is based on a mixture model with constraints on parameters of components’ spread. Experiments with real data show that the proposed approach compares very favorably with other methods of disclosure limitation for continuous microdata in terms of utility and risk.