It would work until algorithms were adjusted to it, which would happen as soon as significant number of people started doing it. Colors are defeated by desaturating, which is no issue since most face recognition algos run on greyscale data anyways. Blotches of bright and dark are defeated, for example, by a high-pass filter (eg: edge detection) on the brightness data to filter out large blotches but keep small detail
back
1 comments
Best bet is an adversarial image, which is essentially reversing what a classification algo is looking for to get out the "Most banana-looking image", for example, and then using that. It, to the AI, looks so much like the category it's designed to look like, that all the weights of everything else in the image are too low in comparison and nothing else gets recognized.