Thread by @conspirator0

If you woke up this morning hoping that somewhere in the world there was a Twitter botnet advertising multiplayer games by replying to tweets (many of them several years old) that have nothing to do with video games, this spam network's for you. #SundaySpam

cc: @ZellaQuixote

This network consists of 45 accounts created between February and June 2020. Almost all of their content (23574 of 24125 tweets, 97.7%) is repetitive replies promoting video games, most of which link to gameexp(dot)com.

At least 23 (probably 25) of the accounts in this network use GAN-generated profile pics. (GAN = "generative adversarial network", the AI technique behind the fake faces produced by thispersondoesnotexist.com etc.) Many have had their colors edited, and 10 have been resized/cropped.

Some of the accounts have profile pics that at first appear to be solid colors. When one increases the level of contrast, two of them (@HannahS21386602 and @RachelA32981013) appear to be faces. We believe these are GAN-generated but due to the lack of detail can't be 100% sure.

Some of the GAN-generated face pics used by this network are used by multiple accounts, with modifications to make them superficially unique.

• @EmmaVanvitelli, @MarinaOvadia, and @SarahBa76005930 have the same pic
• @iXtaktJLC1baCbu and @MaraDoloresMad2 have the same pic

The resizing/cropping throws off techniques for detecting GAN-generated images that depend on the position of the major facial features, but other anomalies remain: some of the resized images feature nonsensical hats/clothing/earrings/glasses, for example.

https://twitter.com/conspirator0/status/1265499541861797890

Here's the result of blending all 25 GAN pics after reversing the modifications to the 10 that had been cropped/resized based on position of the left eye and the distance between eyes, which results in other facial features aligning as they would had the images not been modified.