Re: the hash?
It's not useless because of user behavior. Some users might go through the effort of transforming images until the output has a PhotoDNA vector sufficiently far from the original. Some might even create software to do that automatically. Hell, the really smart ones would train GANs to generate similar images based on an existing corpus – that's a lot cheaper1 than producing real images with real children.
But the vast majority will almost certainly continue to share images verbatim through various off-the-app-store-shelf messaging apps, and for them the PhotoDNA vectors will remain stable.
If the match rate ever fell below whatever target Apple have,2 they can switch to more-sophisticated algorithms. Run each photo through an CNN stack to extract high-level features, compress those, and measure distance in a high-dimensional space, for example. We have lots of classifier architectures that let you split the process at arbitrary points so you're not transmitting the original data.
1Is it less immoral and/or exploitative? There's a fun question for your ethics class.
2And that's probably very, very low. I have no evidence to speculate on how much of this is Apple trying to placate governments and NGOs, and how much is virtue-signalling, and how much is just meant to be the thin edge of the wedge, and how much might even be some sort of misguided altruism. But Apple can't reasonably be expecting to have many true positives, and probably is hoping the positive rate – true and false – is very, very low.
Biting the hand that feeds IT
