Perturbing Nonnormal Confidential Attributes: The Copula Approach

Protecting confidential, numerical data in databases from disclosure is an important issue both for commercial organizations as well as data-gathering and disseminating organizations (such as the Census Bureau). Prior studies have shown that perturbation methods are effective in protecting such confidential data from snoopers. Perturbation methods have to provide legitimate users with accurate (unbiased) information, and also provide adequate security against disclosure of confidential information to snoopers. For databases described by nonnormal multivariate distributions, existing perturbation methods do not provide unbiased characteristics. In this study, we develop a copula-based perturbation method capable of maintaining the marginal distribution of perturbed attributes to be the same before and after perturbation. In addition, this method also preserves the rank order correlation between the confidential and nonconfidential attributes, thereby maintaining monotonic relationships between attributes. The method proposed in this study provides a high level of protection against inferential disclosure. An investigation of the new perturbation method for simulated databases shows that the method performs effectively. The methodology presented in this study represents a signicant step toward improving the practical applicability of data perturbation methods.