Most of the people see the adversary in Apple (or governments), I think there is something else:
What about Adversarial Attacks.
Let’s assume someone is going to spread regular memes modified as Adversarial examples to generate the same neural hash as the true bad images.
Thinking back at the political campaigns, these could spread very easily among some voters for some party.
Suddenly you have a pretty serious attack on people of some political spectrum (or whatever group you can target with this)…
I seriously think this wasn’t fully thought through.
…Maybe we should? All it would take is someone publishing the hash database and then another group working to generate hamming-similar images. Shouldn’t be that hard if it’s as broken as people claim. What’s that saying, “can’t outlaw something that is commonly acceptable behavior”? Point being attacking the system would either prove or compel the false positive rate to be “acceptable”, or result in it being dismantled. Still ideologically opposed but this may be a practical “solution”.
