> The false positive rate for any given image is not 1 in a trillion. Perceptual hashing just does not work like that. It also suffers from the birthday paradox problem - as the database expands, and the total number of pictures expands, collisions become more likely.
There was a good article [0] that was on HN a couple days ago that touches on the flat out lie regarding "one in a trillion" and how PhotoDNA sounds poorly thought out.
Apple is not using PhotoDNA, but in any event - this is a good article but they misconstrued the '1 in a trillion' quote, as is canvassed in the comments for the article itself. According to apple there is a 1 in a trillion chance of wrongly flagging an account, not a 1 in a trillion false positive rate for individual images, and we know that step to banning include:
- multiple flagged images; and
- human review
The details for those processes hasn't been fully disclosed, and it isn't possible to say whether 1 in a trillion is a reasonable estimate or otherwise.
"Don't worry, even if the system flags your baby pictures, we'll just hand them to the lowest-bidding subcontractor who will review the pictures of your naked baby (you know, to catch pedophiles!) and then probably not report you to the authorities. If this happens too often we're just going to ban you tho, this process costs money y'know?"
There was a good article [0] that was on HN a couple days ago that touches on the flat out lie regarding "one in a trillion" and how PhotoDNA sounds poorly thought out.
[0] https://www.hackerfactor.com/blog/index.php?/archives/929-On...