Thread by @conspirator0

Move over, thispersondoesnotexist(dot)com, because deepfake face pics are so 2019. The good folks over on 4chan have made us aware of thiswaifudoesnotexist(dot)net, a site that serves up GAN-generated anime pics. #SaturdayShenaniGANs #ASeriesOfUn4chanateEvents.

cc: @ZellaQuixote

The thiswaifudoesnotexist(dot)net website offers the real anime images used to train the GAN for download (we downloaded via Tor because opsec). We used this along with a set of images it produced to come up with a simple technique for detecting the GAN-generated anime pics.

The GAN-generated anime pics contain a variety of anomalies that aren't present in most of the real ones, but the most obvious (and the one we focused on) is the presence of blotches/very slight random variance in color in areas that would be solid colors on real anime pics.

To detect likely GAN-generated anime pics, we scored the images based on the number of pixels that were barely different from (but not identical to) adjacent pixels. When tested on 5000 real anime images and 5000 GAN-generated ones, it classified 90.2% of the images correctly.

Here's a pastebin link to the code in text format, for those wishing to try it out. There's certainly room for improvement - this is intended as proof-of-concept rather than a polished tool.
pastebin.com/QDKHthRm

Caveats: this technique is intended exclusively to differentiate between real and GAN-generated anime face pics. We do not expect it to produce useful results on other types of images. It is also unlikely to perform well on extremely low-quality JPEG files.

(related thread on detecting the GAN-generated face "photos" produced by thispersondoesnotexist(dot)com. This technique is designed to find most of the GAN-generated pics in a large group of images rather than to classify individual images.)

https://twitter.com/conspirator0/status/1265499437465468935?s=20