[rm999]

tldr: this is expected. The article addresses everything I'm about to say, but I think the lede is buried:

>Apple's NeuralHash perceptual hash function performs its job better than I expected

When you are just looking for any two collisions in 100M x 100M comparisons, of course you'll find a small number of positives, Apple said as much. The number of expected collisions will scale linearly with the number of 'bad' images, which is not 100M. Assuming it's 100k, we'd expect 1000x fewer collisions in ImageNet, or ~0.002 collisions, which is effectively 0. It's the artificial images that will potentially sink all this, not a very low rate of naturally occurring hash collisions.