Comment by odo1242
3 days ago
Yeah, it’s not a great system due to the fact that perceptual hashes can and have been tricked in the past. It is better than machine learning though because you can make any image trigger an ML model without necessarily looking like a bad image. That is, perceptual hashes are much harder to adversarially fool.
I agree, and maybe I'm wrong, but I see a similarity between phash quantization and DCT and ML kernels. I think you could craft "invisible" adversarial images similarly for phash systems like you can ML ones and the results could be just as bad. They'd probably replicate better than adversarial ML images, too.
I think the premise for either system is flawed and both are too error prone for critical applications.