LOCALEVOP-NN: LOCALIZED EVOLVING PRIVATE NEAREST NEIGHBOUR TABULAR DATA SYNTHESIZER

Abstract

Data sharing is important as it helps accumulate novel knowledge, especially in areas like healthcare, finance and social sciences. However, privacy concerns and strict regulations like the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA) lead to a situation where it is quite tricky to access large amounts of that data. In this paper, we have proposed a novel method to produce differentially private tabular synthetic data, named LocalEvoP-NN tabular data synthesizer. LocalEvoP-NN generated synthetic data using private evolution techniques for the k-nearest neighbour. The production of samples is local to the region. The idea of local evolution also reduces the impact of outliers. The method also handles the class imbalance problem by incorporating K-mean SMOTE in the procedure. Our method is successfully applied on four real-world datasets against four different synthesizers, i.e., CTGAN from the SDV package, DPCTGAN, PATECTGAN from the SN-synth package and a non-differential private version of CTABGAN+. LocalEvoP-NN is evaluated on three performance metrics, i.e. AUC/ROC, F1 score and G-mean. Experimentation results show that LocalEvoP-NN outperforms all the other four methods.