None of these cats exist. All are GAN-generated images obtained from thiscatdoesnotexist.com. Can we come up with a way to detect GAN-generated cat pics? #CaturdayShenaniGANs
(GAN = "generative adversarial network", the AI technique used to create the images)
For this project, we used the following datasets (all images 512x512):
• 2000 GAN-generated cats from thiscatdoesnotexist.com
• 1195 real cat faces, cropped from images at kaggle.com/crawford/cat-d…
• a test set of 1000 GAN-generated and 1000 real cats (same sources as above)
Unlike the GAN-generated human face pics provided by thispersondoesnotexist.com etc, the placement of the major facial features on the GAN-generated cat pics from thiscatdoesnotexist.com varies from image to image. There are other anomalies in the fake cat pics, however. . .
We blended 2000 GAN-generated cat pics and altered the color scheme so that each pixel's brightness corresponds to how different that pixel's color is from the average color of the image. This reveals various horizontal and vertical bands in the GAN-generated cat images.
We can emphasize this banding further by blending all the pixels in the same row or column of the image and calculating the difference between adjacent values. This results in a distinctive pattern of peaks every other pixel and slightly less visible peaks every 8 pixels.
The same pattern shows up (albeit less visibly) when we apply the same process to a single GAN-generated cat image. Notably, this pattern is largely absent from real cat face pics, both single and blended.
We used these results to train a simple neural network to classify cat face images as "real" or "GAN-generated". Results are encouraging - within a few hours of training we ended up with a model that achieved 94.2% accuracy on 2000 images it hadn't been trained on.
Here's the code we used for both training and classification. It is presently hardcoded to use 512x512 pixel images (the size generated by thiscatdoesnotexist.com) and will likely require tweaking for other image sizes. pastebin.com/hqPtD5CD
A couple of disclaimers:
• This method is unlikely to produce useful results on images that aren't of cat faces.
• If using it on Twitter profile pics, you will want to obtain the full-resolution version of the pics for best results (more detail here: developer.twitter.com/en/docs/twitte…).
• • •
Missing some Tweet in this thread? You can try to
force a refresh
• Community Notes successfully placed fact checks on some of the most viral false posts about the shooting
• ~42% of noted posts were subsequently deleted by their authors
• An effort to spread a misidentification of the shooter via Community Notes failed
THE BAD:
• Community Notes fact checks take several hours to show up, which doesn't help much in the initial "breaking news" phase after a violent event
• Many notes never accumulate enough ratings to determine their fate
12 questions for @TheDailyBeast regarding @JakeLahut's false April 2023 story, "How Ron DeSantis Is Taking a Page Out of Nixon’s Playbook", which (among other things) falsely portrays an AI-generated face as a "sexually graphic meme" of a real child.
@JoannaColes @TracyConnor
First, some background and a couple debunks of the false article, for those unfamiliar with the situation:
1. How did the decision to use serial fabulist Steven Jarvis as a source for this article come about?
2. Was anyone employed by or affiliated with The Daily Beast at the time the article was published aware of Steven Jarvis's extensive history of making false claims?
Meet @LovewinnLove (permanent ID 2707213009), a blue-check verified account with a GAN-generated face and a few additional odd characteristics. Despite being created in 2014, this account has no posts prior to October 2023.
cc: @ZellaQuixote
There are multiple indicators that @LovewinnLove's "face" is GAN-generated:
• unrealistic teeth (visible portion of bottom teeth is especially bizarre)
• odd texturing and seams in shirt fabric
• telltale eye positioning (more info in next post)
@LovewinnLove All unmodified StyleGAN-generated face images have the property that the major facial features (particularly the eyes) are in the same position on each image. Blending @LovewinnLove's profile image with 99 other GAN-generated faces demonstrates this nicely.
It's New Year's Eve, and a bunch of politics enthusiasts with GAN-generated faces are enthusiastically replying to a variety of posts with similarly-worded replies. #NewYearShenaniGANs
cc: @ZellaQuixote
The politics enthusiasts are part of a spam network consisting of (at least) 575 accounts created between May and December 2023 with GAN-generated faces. Many of their handles, such as @Maairiuieinaaa and @eJooeiaAoneueer, contain long strings of vowels.
@Maairiuieinaaa @eJooeiaAoneueer All 575 of these accounts use StyleGAN-generated faces as profile images. Some of these, such as @MauMoiagaia's profile image, contain a tiny "StyleGAN 2 (Karras et al.)" watermark in the lower right corner.
It's a great day to look at a network of inauthentic accounts that post identical AI art images (with a side of good old fashioned T-shirt spam).
cc: @ZellaQuixote
This network consists of 24 X accounts. 12 of these accounts were created in the latter half of 2023 and have female avatars, while the other 12 were created in 2013 or earlier and have male avatars.
The 12 accounts with female avatars and 2023 creation dates regularly post AI-generated art images, and these image posts are quickly reposted by other accounts in the network (both female and male). The AI-generated images are often duplicated across accounts.
Meet @ImJamesMiller (permanent ID 1371651462153994242), an account with a GAN-generated face, 172K followers, and no tweets prior to two days ago. What's up with that?
cc: @ZellaQuixote
As it turns out, @ImJamesMiller wasn't always named @ImJamesMiller. In June, the account was named @/IamJimCaviezel in an apparent attempt to impersonate Sound of Freedom actor Jim Caviezel.
@ImJamesMiller Multiple prominent users appear to have accepted the fake Jim Caviezel account as legitimate, including Texas Congressman Brian Babin, right-wing influencer/ex-Game of Thrones blogger Jack Posobiec, and recently indicted ex-Assistant Attorney General Jeff Clark.