DT-Anon: Decision Tree Target-Driven Anonymization

More and more scenarios rely today on data analysis of massive amount of data, possibly contributed from multiple parties (data controllers). Data may, however, contain information that is sensitive or that should be protected (e.g., since it exposes identities of the data subjects) and cannot simply be freely shared and used for analysis. Business rules, restrictions from individuals (data subjects to which data refer), as well as privacy regulations demand data to be sanitized before being released or shared with others. Unfortunately, such protection typically comes with a loss of utility of the released data, impacting the performance of the analytics tasks to be executed. In this paper, we present DT-Anon, a target-driven anonymization approach that aims at protecting (anonymizing) data while preserving as much as possible the capability of a classification task operating downstream to learn from the anonymized data. The basic idea of our approach is to perform the anonymization process on partitions produced by a decision tree driven by the target of the classification task. Each partition is then independently anonymized, to limit the impact of anonymization on the attributes and values that work as predictors for the target of the classification task. Our experimental evaluation confirms the effectiveness of the approach.