[ZephyrBlu]

I doubt this is true. I've seen adversarial version of images which are very similar to the original image.

E.g. an image of a cat that has had some pixels flipped so it's classified as a dog instead.

[haha69]

Something like flipping a pixel would work if you manipulate the actual image (raw image) being fed to the software.

Here it is captured via a camera and then fed to the software and I doubt pixel level details are captured by a camera

[ZephyrBlu]

I don't mean literally flipping pixels, but subtle changes in the image that cause a misclassification. See this post for examples: https://pyimagesearch.com/2020/10/19/adversarial-images-and-....

[haha69]

Yes, I had a sense for what you were referring to and my reply was geared towards that. Not just flipping pixels.

[ZephyrBlu]

Right, I understand what you mean now. You're saying that those subtle changes might not be captured by the camera, which makes sense. I don't have any experience with CCTV systems.

[dylan604]

okay, but i'm not going in for an elective surgery to have some of the pixels on my actual face flipped. i'd rather just wear a hat and other things to disrupt/block cameras from seeing my face

[ZephyrBlu]

You misunderstand. The clothing on the linked website does not cover your face either. A fairly normal looking piece of clothing specifically designed to exploit the classification model could likely be used to "confuse" facial recognition rather than needing a piece of clothing with a garish design.

[dylan604]

No I didn't misunderstand. I'm well aware of what clothing is and how it is used.

You described an adversarial image as one that has pixels flipped. That has nothing to do with clothing either, and had nothing to do with real time facial recognition as this "adversarial" clothing is meant to disrupt. So I just took your meaningless pixel flipping suggestion back to subject at hand.

Also, from the examples I've seen, facial recognition has no problems recognizing multiple faces in the same image. So I just don't understand the point of clothing like this when all it is going to do is present the software with a few additional things to consider, but not actually stop it from consider the actual face of the wearer of the clothing.