> attackers also put a small red circle in the middle of images to throw off computer vision detection tools.
I didn't get this. On Googling, I think it is mis-reported.
One Pixel Attack for Fooling Deep Neural Networks, Su et al
Putting ONE wrong pixel at a selected point in an image can make computer vision detection tools see a ship as a car, a (upside down?) car as an airplane, a cat or horse as a dog.
The "red circle" appears on page 2 of the paper, to highlight the one wrong pixel for reader reference. This would not be the representation on the webpage.
I suppose if a page instructs "Click on the dog" and get a click on the cat, it may presume the clicker is a 'bot and should not be phished.
But some of these images are as horrible as the Captchas on 4Chan. What the heck IS that?