An Empirical Study of Membership Inference Attacks Against Tabular Data Synthesis Models

Experiments on 4 state-of-the-art tabular data synthesis models found that membership inference attacks can determine whether a given record was used to train a victim synthesis model and can severely compromise these models. The experiments cover both black-box and white-box attack scenarios. The study also evaluates the protective effect of two differential privacy training algorithms, DP-SGD and DP-GAN, with the key finding that both can substantially mitigate this threat, but at the cost of sacrificing generation quality.