ИСТИНА

Generative Adversarial Network (GAN) has proven to be a highly efficient technology. It shows remarkable results in a wide range of applications. However, sufficient and representative datasets are prerequisite for successful training of GAN. These datasets are often crowdsourced and outsourced and may contain private information. These models must not expose private information because disclosure of such data is treated as a breach of law. In this research study, we investigate attacks against generative models proposed in [1]: given a record, the adversary determines if it was present in the training dataset. These attacks are aimed at detecting overfitting of GAN, which might lead to a disclosure of data. We have discovered that even a simplistic GAN can be potent to secure private information if one follows recommendations given in [1].