This network consists of 24 accounts created between May 2019 and December 2020. All have GAN-generated face images as their profile pics. Presently, all 24 (allegedly) tweet via the Twitter Web App.
The current generation of GAN-generated face pics have the anomaly that the major facial features (particularly the eyes) are in the same pixel position on each image. This trait becomes easy to see when we blend the images together, as in this video:
This network's content is mostly retweets, and the accounts mostly retweet one another - 26806 of 33786 retweets (79.3%) are retweets of other members of the network.
The network's original content is mostly hashtag spam, occasionally containing links to blog posts on digital marketing/SEO ("search engine optimization"). The hashtags used are marketing/SEO related, with #SEO, #DigitalMarketing, #SMM, and #Marketing, and being the most common.
In addition to retweeting one another, the accounts in this network also follow each other profusely: each member of the network follows an average of 19 of the other 23 accounts in the network.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
This video shows the process of blending @JaredLCarter's profile pic with 9 pictures generated by thispersondoesnotexist.com, demonstrating that the major facial features (particularly the eyes) are in the exact same place, a fingerprint of unmodified GAN-generated face pics.
(more threads on the use of GAN-generated images and how to detect them here:
One of the accounts that @DrunkAlexJones was forced to follow by the "Round Year Fun" app is @drago1171, an account created in 2009 with a GAN-generated face pic that follows several other old accounts with GAN-generated face pics. #SaturdayShenaniGANs
By recursively exploring the followers/followees of these accounts, we found a network of 54 accounts created in 2009 or 2010 with GAN-generated faces. All were dormant from late 2012 until April 2021, when all 54 began tweeting via the Twitter Web App.
Here's an animated visualization of blending all 54 profile pics together, showing the identical facial feature placement that is a fingerprint of GAN-generated pics such as those created by thispersondoesnotexist.com. (GAN = "generative adversarial network", the AI technique used.)
PSA: These "fun" apps from roundyear(dot)fun ("My Twitter Family" etc) have a downside: they gain near-total control of your account and (at the very least) use it to follow other accounts without your knowledge. #FunAllYearRoundUntilYourAccountGetsCompromised
We had @DrunkAlexJones test some of the Round Year Fun apps. The list of permissions the apps request is extensive and encompasses pretty much every action one could possibly take with one's Twitter account. The apps produced the expected "My Twitter Crush" etc tweets.
These eight accounts were all created on either February 18th or February 21st, 2021. Thus far, they have tweeted very little (only 58 tweets), and have sent all their tweets via the Twitter Android app. Two of them (@An_mal12 and @An_mal14) currently have identical biographies.
The majority of these accounts' tweets so far are extremely brief replies to each other (and sometimes replies to each other's replies to each other, and so on). They have also posted a few follower growth tweets, including retweets of a tweet that contains a likely malware link.
This botnet is comprised of 4959 accounts created between January 7th and January 28th, 2016 (with a mid-month break). All of them follow between 35 and 40 accounts, and have never liked a tweet. Account names consist of English first names followed by four random letters.
This network only uses 23 unique profile pictures (22 photos and the default pic) across 4959 accounts, resulting in each non-default pic appearing on over 200 accounts. Reverse image searches indicate that the same images may have been used for spam networks on Facebook as well.
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. . .