Summary: Lately, generative adversarial networks (GANs) have achieved beautiful realism, fooling even human observers. Certainly, the favored tongue-in-cheek web site http://thispersondoesnotexist.com, taunts customers with GAN generated pictures that appear too actual to consider. Alternatively, GANs do leak details about their coaching knowledge, as evidenced by membership assaults lately demonstrated within the literature. On this work, we problem the idea that GAN faces actually are novel creations, by setting up a profitable membership assault of a brand new sort. In contrast to earlier works, our assault can precisely discern samples sharing the identical identification as coaching samples with out being the identical samples. We display the curiosity of our assault throughout a number of standard face datasets and GAN coaching procedures. Notably, we present that even within the presence of great dataset variety, an over represented individual can pose a privateness concern.