We've done a decent amount of research on the use of GAN-generated images over the last two years, mostly fake face "photos" such as those produced by thispersondoesnotexist(dot)com. Here are all of our related threads in one place.
GAN is an abbreviation for "generative adversarial network", the AI technique used to produce these images (and other things). Here's a "brief" description of how StyleGAN, the GAN behind thispersondoesnotexist(dot)com, works. machinelearningmastery.com/introduction-t…
Here's the first network we found in the wild using GAN-generated face pics (29 of 52 accounts, created in two batches). The profiles featured repetitive biographies mentioning family, friends, country, and weapons.
In May 2020, we ran into a set of five fake pro-Biden accounts, all using the same GAN-generated profile pic. These accounts self-deleted shortly after this thread was posted.
Our first stab at a programmatic method for detecting GAN-generated face pics. This technique is designed to make the GAN-generated pictures easier to notice in groups of thousands of images - it isn't terribly accurate for individual images.
Next, a 4chan op using accounts with GAN-generated pics to push the #BernReturn hashtag alongside a bogus claim that Bernie Sanders was returning campaign donations.
Speaking of 4chan, thispersondoesnotexist(dot)com is quite popular over there, and folks frequently suggest using it as a source of profile pics for bogus Twitter accounts.
Here's a thread on now-suspended part-automated political troll account @Juan6million. This account mostly pushed left-wing messaging and hashtags, but also promoted alt-right talking points and influencers such as @MrAndyNgo.
The most prolific use of GAN-generated profile pics we've seen thus far was from the "Thousand Followers" follower-buying website (presently on hiatus). 8727 of 17957 of the fake Twitter followers provided by this site used GAN-generated images.
We've run across a couple of botnets promoting cryptocurrency sites that use GAN-generated profile pics: 41 accounts pushing cointelegraph(dot)com and 28 accounts pushing ethereumcryptocurrency(dot)com.
Deepfake human faces are not the only sort of images one can produce with GANs. Here's a thread on GAN-generated anime pics, and the detection thereof:
A similar network follower/retweet botnet created in August 2020. 53 accounts, all using GAN-generated pics. It mostly amplifies Russian-language accounts.
Thread on a weird automated account (@Jtatejtate1) using a GAN-generated face. A bunch of its followers are part of a recently reactivated botnet from 2013.
Some thoughts on perennial pitfalls in news coverage of social media manipulation that frequently result in reporting on fake accounts/bots/etc being far less accurate and informative than it ought to be...
The most common problem with news articles about fake accounts: failure to include any examples of fake accounts or evidence of their inauthenticity. Any or all of these headlines might be accurate, but you can't tell from the articles, due to absence of evidence.
A related issue: articles like the "Nearly Half of Biden/Trump's Followers Are Fake" and "Nearly Half Of Accounts Tweeting About Coronavirus Are Bots" pieces base their numbers on closed-source third party tools, which may or may not actually be detecting anything useful.
Does thanking, praising, or insulting an LLM-based chatbot affect the speed or accuracy of its responses to questions involving basic arithmetic? Let's find out!
For this experiment, Meta’s Llama 3.1 model was asked to add and multiply random numbers between 10 and 100, with six different wordings: polite, rude, obsequious, urgent, and short and long neutral forms. Each combination of math operation and wording was tested 1000 times.
Results: asking the questions neutrally yielded a faster response than asking politely, rudely, obsequiously, or urgently, even if the neutral prompt was longer. Overall, obsequious math questions took the longest to process, followed by urgent, rude, and polite questions.
Just for fun, I decided to search Amazon for books about cryptocurrency a couple days ago. The first result that popped up was a sponsored listing for a book series by an "author" with a GAN-generated face, "Scott Jenkins".
cc: @ZellaQuixote
Alleged author "Scott Jenkins" is allegedly published by publishing company Tigress Publishing, which also publishes two other authors with GAN-generated faces, "Morgan Reid" and "Susan Jeffries". (A fourth author uses a photo of unknown origin.)
As is the case with all unmodified StyleGAN-generated faces, the facial feature positioning is extremely consistent between the three alleged author images. This becomes obvious when the images are blended together.
The people in these Facebook posts have been carving intricate wooden sculptures and baking massive loaves of bread shaped like bunnies, but nobody appreciates their work. That's not surprising, since both the "people" and their "work" are AI-generated images.
cc: @ZellaQuixote
In the last several days, Facebook's algorithm has served me posts of this sort from 18 different accounts that recycle many of the same AI-generated images. Six of these accounts have been renamed at least once.
The AI-generated images posted by these accounts include the aforementioned sculptures, sad birthdays, soldiers holding up cardboard signs with spelling errors, and farm scenes.
The common element: some sort of emotional appeal to real humans viewing the content.
As Bluesky approaches 30 million users, people who run spam-for-hire operations are taking note. Here's a look at a network of fake Bluesky accounts associated with a spam operation that provides fake followers for multiple platforms.
cc: @ZellaQuixote
This fake follower network consists of 8070 Bluesky accounts created between Nov 30 and Dec 30, 2024. None has posted, although some have reposted here and there. Almost all of their biographies are in Portuguese, with the exception of a few whose biographies only contain emoji.
The accounts in this fake follower network use a variety of repeated or otherwise formulaic biographies, some of which are repeated dozens or hundred of times. Some of the biographies begin with unnecessary leading commas, and a few consist entirely of punctuation.
It's presently unclear why, but over the past year someone has created a network of fake Facebook accounts pretending to be employees of the Los Angeles Dodgers. Many of the accounts in this network have GAN-generated faces.
cc: @ZellaQuixote
This network consists of (at least) 80 Facebook accounts, 48 of which use StyleGAN-generated faces as profile images. The remaining 32 all use the same image, a real photograph of a random person sitting in an office.
As is the case with all unmodified StyleGAN-generated faces, the main facial features (especially the eyes) are in the same position on all 48 AI-generated faces used by the network. This anomaly becomes obvious when the faces are blended together.