47d1e990583c9c67424d369f3414728e-AuthorFeedback.pdf

Wewillhighlight7 theseresultsintherevision.8 It is better to demonstrate some evaluations of the information leakage of Bayesian GAN. Although the pure theoretical13 part is somewhat straightforward, it demonstrates a novel and practical view of privacy protection to analyze the14 generalization property. This21 empirical evidence indicates that we can improve the generalization ability of GANs by reducing the individual's22 information leakage. The results38 of the black-box attack (in Table 2) also show that Lipschitz constraints can reduce the information leakage of the39 generator (since the black-box attack depends on the information of the generator). This is meaningful in some40 real-world applications, such as releasing the learned generator to the public for the research purpose.