Pufferfish Privacy Mechanism Based on Multi-dimensional Markov Chain Model for Correlated Categorical Data Sequences

Abstract

Differential privacy is a rigorous standard for protecting data privacy and has been extensively used in data publishing and data mining. However, because of its vulnerable assumption that tuples in the database are in-dependent, it cannot guarantee privacy if the data are correlated. Kifer et al. proposed the Pufferfish Privacy framework to protect correlated data privacy, while till now under this framework there is only some practical mechanism for protecting correlations among attributes of one individual sequence. In this paper, we extend this framework to the cases of multiple correlated sequences, in which we protect correlations among individual records, as well as correlations of attributes. Application scenarios can be different people’s time-series data and the objective is to protect each individual’s privacy while publishing useful information. We firstly define privacy based on Pufferfish privacy framework in our application, and when the data are correlated, the privacy level can be assessed through the framework. Then we present a multi-dimensional Markov Chain model, which can be used to accurately describe the structure of multi-dimensional data correlations. We also propose a mechanism to implement the privacy framework, and finally conduct experiments to demonstrate that our mechanism achieves both high utility and privacy.