Progress Made by Our Center in Synthetic Data Privacy Leakage Risk Assessment Technology
In recent years, generative adversarial network (GAN) techniques have been widely adopted for time-series data synthesis to alleviate data scarcity and address privacy compliance issues. As a representative approach in time-series data synthesis, TimeGAN integrates self-supervised learning and adversarial training, demonstrating superior performance in preserving statistical properties and temporal dependencies. Nevertheless, whether TimeGAN is genuinely "secure" and whether its training and generation processes entail privacy leakage risks remain under-investigated in a systematic manner.
To tackle this critical issue, the research team from the Security Department of our center conducted the first systematic evaluation of the privacy leakage risks of TimeGAN under membership inference attacks (MIA). From the perspective of attackers’ knowledge, this study established two types of privacy attack scenarios, i.e., white-box and black-box settings, to comprehensively characterize the potential leakage risks faced by TimeGAN in various application environments. The findings can provide a full-process and complete technical toolkit for institutional privacy risk self-assessment (white-box) and privacy auditing (black-box), while laying a theoretical and methodological foundation for the subsequent development of differential privacy protection mechanisms tailored for time-series data generation models.
Privacy Leakage Risk Assessment Flowchart
The research results have been accepted by ICASSP 2026, a authoritative conference in the field of signal processing and audio technology recommended as a CCF Class B conference. The first author of the paper is Zhang Ninghui, a master’s student from the Security Department of our center, and the corresponding author is Senior Engineer Long Chun. This research was supported by the National Key R&D Program of China (Grant No. 2023YFC3304704).
Related Publication:Zhang N, Long C, Li J, et al. Uncovering Privacy Risks in TimeGAN: Novel and Effective Membership Inference Attacks[C]//ICASSP 2026.