Time-dependent differential privacy for enhanced data protection in synthetic transaction generation
Kirill Mikhailovich Zakharov, Elizaveta Stavinova · 2024
In today’s data-driven world, the increasing concerns about data privacy and security have made privacy-preserving methods a necessity to protect sensitive information. In this paper, we extend our TRGAN method [25] by privacy-preserving mechanisms within GAN-based synthetic transaction generation. We introduced four mechanisms, one of which is based on a novel approach called time-dependent differential privacy. Additionally, we provide a stochastic implementation of this mechanism, allowing to identify the privacy budget according to the generated data. Our experiments show that (1) the robustness of transaction generation results with time-dependent differential privacy has improved in comparison with existing privacy-preserving modifications; (2) the stochastic implementation of time-dependent differential privacy provides a more effective way of generating varying levels of noise in privacy mechanisms.