In a series of experiments published in Nature Communications, we found evidence that human judgments are indeed systematically influenced by adversarial perturbations.
Disclaimer: I have no real qualification on this. But it seems like this whole technology is pretty sensitive to the specific model being used and the specific details of the pixels; the whole thing is written like there’s some silver-bullet image alteration that can fool “machine vision” in general, but what it demonstrates is nothing like that.
I asked Midjourney to identify the altered images that machines are supposed to identify as a sheep or a cat or whatever, and it said:
A bouquet of flowers sitting on the table in a brown vase
Some bright colored flowers in a circular vase
An omelette and sandwiches on the table
An omelet with hash browns
… which is what they are.
The last two images were actually a little more interesting – they’re distorted to the point that it’s visually obvious that they’ve been altered, and Midjourney actually picks up that the image is distorted a little, and includes that in the style part of its description, while mostly-accurately describing what’s in the image. These are its full descriptions:
“a red bridge, traffic lights, and a fencedin section of street, in the style of digital mixed media, thermal camera, american realism, found object sculpture, stipple, ricoh r1, xbox 360 graphics”
“a pole with a traffic light and a van, in the style of distorted, fragmented images, manapunk, found objects, webcam photography, suburban ennui capturer, hyper-realistic bird studies, 19th century american art”
Disclaimer: I have no real qualification on this. But it seems like this whole technology is pretty sensitive to the specific model being used and the specific details of the pixels; the whole thing is written like there’s some silver-bullet image alteration that can fool “machine vision” in general, but what it demonstrates is nothing like that.
I asked Midjourney to identify the altered images that machines are supposed to identify as a sheep or a cat or whatever, and it said:
… which is what they are.
The last two images were actually a little more interesting – they’re distorted to the point that it’s visually obvious that they’ve been altered, and Midjourney actually picks up that the image is distorted a little, and includes that in the style part of its description, while mostly-accurately describing what’s in the image. These are its full descriptions:
“a red bridge, traffic lights, and a fencedin section of street, in the style of digital mixed media, thermal camera, american realism, found object sculpture, stipple, ricoh r1, xbox 360 graphics”
“a pole with a traffic light and a van, in the style of distorted, fragmented images, manapunk, found objects, webcam photography, suburban ennui capturer, hyper-realistic bird studies, 19th century american art”