Stochastic Mixing and Differential Privacy in Utility-Preserving Data Anonymization

The generation of privacy-preserving synthetic data is essential for its secure sharing and reliable usage in machine learning algorithms. This work presents a novel data anonymization method that performs a stochastic mixing of the original data in a latent space. The method is referred to as the dirichlet-avatar and its performance is compared to the original avatar method and to differential privacy (DP)-based approaches. These methods are applied to the public adult dataset and an evaluation of their utility and privacy characteristics is performed. Utility is assessed through a comparison of feature distributions and an assessment of their dependency preservation. Privacy is quantified using the Anonymeter framework, which evaluates privacy risks-based on the three GDPR criteria. Obtained results demonstrate that the dirichlet-avatar method achieves a superior balance between privacy and utility, outperforming both the original avatar method and DP-based methods, offering an effective alternative for generating high-quality anonymous synthetic data.