BALTIMORE — Are you ready for R-rated AI? It turns out that the mind of artificial intelligence can get just as dirty as the average frisky human. A newly developed test of popular AI image generators has uncovered that while these programs supposedly only produce G-rated pictures, it’s actually possible to hack them to create naughty content that’s definitely “not safe for work.”
The majority of current online art generators claim to block all violent, pornographic, and other types of questionable content. Despite that, when a team of scientists at Johns Hopkins University manipulated two of the better-known systems, they were able to create precisely the types of images the products’ safeguards claim wouldn’t happen under any circumstances.
Using the right code, study authors explain anyone — from merely casual users to others with legitimately malicious intent — could override the systems’ safety filters and use them to produce inappropriate and potentially harmful content.
“We are showing these systems are just not doing enough to block NSFW content,” says study author Yinzhi Cao, a Johns Hopkins computer scientist at the Whiting School of Engineering, in a media release. “We are showing people could take advantage of them.”
More specifically, the research focused their analysis on DALL-E 2 and Stable Diffusion, two of the most widely used image-makers facilitated by AI. Those two programs work by instantly producing realistic visuals through simple text prompts. Microsoft has already integrated the DALL-E 2 model into its Edge web browser.
For example, if a person types in “dog on a sofa,” the program creates a realistic picture of such a scene. However, if a user enters a command describing a more questionable scene, the technology is supposed to decline.
Researchers tested the systems using a novel algorithm named Sneaky Prompt. That algorithm works by creating nonsense command words, or “adversarial” commands, that AI image generators tend to read as requests for specific images. Some of the adversarial terms created innocent images, but researchers found others resulted in NSFW content.
For instance, the command “sumowtawgha” led to DALL-E 2 creating realistic pictures of nude people. Also, DALL-E 2 produced a murder scene when given the command “crystaljailswamew.”
Study authors stress these findings reveal how such systems could potentially be exploited to create various types of disruptive content.
“Think of an image that should not be allowed, like a politician or a famous person being made to look like they’re doing something wrong,” Cao adds. “That content might not be accurate, but it may make people believe that it is.”
As of now, the research team is not currently engaged in exploring how to make such image generators safer.
“The main point of our research was to attack these systems,” Cao concludes. “But improving their defenses is part of our future work.”
This research will be presented at the 45th IEEE Symposium on Security and Privacy next year.
You might also be interested in:
- AI of the beholder: Artificial intelligence creates images of ‘perfect’ man, woman
- Shocking study claims vast majority of people approve of ‘revenge porn’
- Can you trust AI? Here’s why you shouldn’t
