OpenAI researchers believe they have discovered a shockingly easy way to hoodwink their object-recognition software, and it requires just pen and paper to carry out.
Specifically, the lab's latest computer vision model, CLIP, can be tricked by in what's described as a “typographical attack." Simply write the words ‘iPod’ or ‘pizza’ on a bit of paper, stick it on an apple, and the software will wrongly classify the piece of fruit as a Cupertino music player or a delicious dish.
“We believe attacks such as those described above are far from simply an academic concern,” the bods behind CLIP said this week. “By exploiting the model’s ability to read text robustly, we find that even photographs of hand-written text can often fool the model." They added that "this attack works in the wild," and "it requires no more technology than pen and paper.”
CLIP isn't the only artificially intelligent software to fall for such simple shenanigans. It was demonstrated you could use sticky tape to fool Tesla’s Autopilot into misreading a 35mph sign as an 85mph one. Other forms of these so-called adversarial attacks, however, require some technical know-how to execute: it typically involves adding noise to a photo or crafting a sticker of carefully arranged pixels to make an object-recognition system mistake, say, a banana for a toaster. In CLIP’s case, however, none of that is necessary.
Suffice to say, OpenAI's model was trained using pictures of text as well as images of objects and other things scraped from the internet.
Think your smartwatch is good for warning of a heart attack? Turns out it's surprisingly easy to fool its AIREAD MORE
This approach was taken so that CLIP remains fairly general purpose, and can be fine-tuned as needed for a particular workload without having to be retrained. Given an image, it can not only predict the right set of text labels describing the scene, it can be repurposed to search through large databases of pictures and provide captions.
CLIP is able to learn abstract concepts across different representations, OpenAI said. For example, the model is able to recognize Spider-Man when the superhero is depicted in a photo, a sketch, or described in text. What’s more interesting is that the researchers have been able to find groups of neurons in the neural network that are activated when the software clocks a glimpse of Spider-Man.
They have described these as multimodal neurons. “One such neuron, for example, is a 'Spider-Man' neuron that responds to an image of a spider, an image of the text 'spider,' and the comic book character 'Spider-Man' either in costume or illustrated,” the OpenAI team said. CLIP has all sorts of multimodal neurons that represent different concepts, such as seasons, countries, emotions, and objects.
But the model’s greatest strengths – its versatility and robustness – is also its greatest weakness. CLIP is easily hoodwinked by typographical attacks, they found.
Object-recognition AI – the dumb program's idea of a smart program: How neural nets are really just looking at texturesREAD MORE
Going back to the apple vs pizza example, the multimodal neurons that have learnt the representation of an apple don’t fire as well when they see the written word 'pizza.' Instead, the pizza-related neurons get triggered instead. The model is easily confused.
There is evidence that abstract learning using multimodal neurons also occurs in human brains. But unfortunately, here’s where modern machines pale in comparison to their biological counterparts. Humans can obviously tell that an apple with a handwritten note that reads pizza on it is still an apple, while AI models can’t yet.
OpenAI said CLIP doesn't perform as well as some computer vision models that are today used in production. It also suffers from offensive biases, its neurons associate the concept of the 'Middle East' with 'terrorism' and black people with gorillas. The model is only used for research purposes at the moment, and OpenAI is still deciding whether or not to release the code.
"Our own understanding of CLIP is still evolving, and we are still determining if and how we would release large versions of CLIP. We hope that further community exploration of the released versions as well as the tools we are announcing today will help advance general understanding of multimodal systems, as well as inform our own decision-making," it said.
OpenAI declined to comment further on CLIP. ®