But the remarkable thing there (and with all other adversarial attacks I've seen) is that the ML classifier is fooled, while for us humans it is obvious that it is still the original image (if maybe slightly perturbed).

But in the case of Apple's CSAM detection, the collision would first have to fool the victim into seeing an innocent picture and storing it (presumably, they would not accept and store actual CSAM [^]), then fool the NeuralHash into thinking it was CSAM (ok, maybe possible, though classifiers <> perceptual hash), then fool the human reviewer into also seeing CSAM (unlike the innocent victim).

[^] If the premise is that the "innocent victim" would accept CSAM, then you might as well just send CSAM as an unscrupulous attacker.

Hmm, not quite:

step 1 - As others have pointed out, there are plenty of ways of getting an image onto someone's phone without their explicit permission. WhatsApp (and I believe Messenger) do this by default; if someone sends you an image, it goes onto your phone and gets uploaded to iCloud.

step 2 - TFA proves that hash collision works, and fooling perceptual algorithms is already a known thing. This whole automatic screening process is known to be vulnerable already.

step 3 - Humans are harder to fool, but tech giants are not great at scaling human intervention; their tendency is to only use humans for exceptions because humans are expensive and unreliable. This is going to be a lowest-cost-bidder developing-country thing where the screeners are targeted on screening X images per hour, for a value of X that allows very little diligence. And the consequences of a false positive are probably going to be minimal - the screeners will be monitored for individual positive/negative rates, but that's about it. We've seen how this plays out for YouTube copyright claims, Google account cancellations, App store delistings, etc.

People's lives are going to be ruined because of this tech. I understand that children's lives are already being ruined because of abuse, but I don't see that this tech is going to reduce that problem. If anything it will increase it (because new pictures of child abuse won't be on the hash database).

> then fool the human reviewer into also seeing CSAM (unlike the innocent victim).

Or just blackmail and/or bribe the reviewers. Presumably you could add some sort of 'watermark' that would be obvious to compromised reviewers. "There's $1000 in it for you if you click 'yes' any time you see this watermark. Be a shame if something happened to your mum."

Yes but the reviewers are not going to be viewing the original image, they are going to be viewing a 100x100 greyscale.

>If the premise is that the "innocent victim" would accept CSAM, then you might as well just send CSAM as an unscrupulous attacker.

This adds trojan horses embedded in .jpg files as an attack vector, which while maybe not overly practical, I could certainly imagine some malicious troll uploading "CSAM" to some pornsite.

NN classifiers work differently than perceptual hashes and the mechanism to do this sort of attack is entirely different, though they seem superficially similar.