A new approach for dynamic and risk-based data anonymization

Abstract

Data anonymization is a complex task, as it is dependent on the structure of the dataset, the privacy requirements that we might have and how the anonymized data is going to be processed. Taking just into account these three aspects, it would be possible to set up many anonymization configurations for a single dataset, as each variable that appears on the data could be anonymized using different techniques (generalization, randomization, deletion), and each of them could be configured with a different parameterization. In consequence, the are several alternatives for anonymizing a dataset, especially when it is composed by a high number of variables. For those cases, a manual anonymization process is unfeasible and an automatic approach that allows to determine the best anonymization configuration for the data is essential. Furthermore, it is necessary to determine accurately the risk of each anonymization configuration, in order to verify that the expected privacy requirements are fulfilled. In this paper we present two main contributions: 1) a dynamic risk-based anonymization process that allows to determine the best anonymization configuration for a particular dataset; 2) two new privacy metrics (CAK and R-CAK) that allow to measure the risk of re-identification of the anonymized data, taking into account the knowledge of an adversary that is trying to disclose sensitive attributes from the anonymized dataset.