Sensitive Label Privacy Preservation with Anatomization for Data Publishing

摘要

Data in its original form, however, typically contain sensitive information about individuals. Directly publishing raw data will violate the privacy of people involed. Consequently, it becomes increasingly important to preserve the privacy of published data. An attacker is apt to identify an individual from the published tables, with attacks through the record linkage, attribute linkage, table linkage or probabilistic attack. Although algorithms based on generalization and suppression have been proposed to protect the sensitive attributes and resist these multiple types of attacks, they often suffer from large information loss by replacing specific values with more general ones. Alternatively, anatomization and permutation operations can de-link the relation between attributes without modifying them. In this paper, we propose a scheme Sensitive Label Privacy Preservation with Anatomization (SLPPA) to protect the privacy of published data. SLPPA includes two procedures, table division and group division. During the table division, we adopt entropy and mean-square contingency coefficient to partition attributes into separate tables to inject uncertainty for reconstructing the original table. During the group division, all the individuals in the original table are partitioned into non-overlapping groups so that the published data satisfies the pre-defined privacy requirements of our (alpha,beta,gamma,delta) model. Two comprehensive sets of real-world relationship data are applied to evaluate the performance of our anonymization approach. Simulations and privacy analysis show our scheme possesses better privacy while ensuring higher utility.