Statistical Distribution-Aware and Privacy-Preserving Anonymization using Generative Adversarial Networks

The growing availability of tabular data in healthcare, public administration, and socio-economic studies presents significant privacy challenges. Traditional methods such as k-anonymity remain vulnerable to linkage attacks. Generative models, particularly CTGAN, enable the production of realistic synthetic data while preserving multivariate dependencies. However, existing DP-CTGAN models apply differential privacy uniformly, which can degrade data utility by disrupting the distributions of non-sensitive attributes. We propose a semi-synthetic approach in which DP is applied only to sensitive or potentially re-identifiable attributes via the discriminator, while non-sensitive attributes are preserved. This strategy maintains the statistical stability of non-sensitive attributes while protecting sensitive information. On the Adult dataset, as well as on a generated socio-economic and healthcare dataset of 30,000 records, evaluated using the NNAA metric, our method demonstrates an improved privacy–utility trade-of, preserving fine correlations and high statistical fidelity.