3 ms·

Arent the perceptual hashes based on a chunk of the image? I wonder what the false positive rates are for: - A random image against the DB of perceptual hashe

by cwkoss 5y ago

Arent the perceptual hashes based on a chunk of the image? I wonder what the false positive rates are for: - A random image against the DB of perceptual hashes - Images of a baby's skin against the DB of perceptual hashes It seems like the second would necessarily have a higher false positive rate: similar compositions (contains baby's skin) would more likely have similar chunks. Is it just a little higher or several orders of magnitude higher? I know hash collisions are rare, but wonder how much rarity of collisions decreases with perceptual hashes.

fossuser 5y ago
It's two factors, both the match on an image hash and an unknown threshold of matches at which point the data gets sent up. If the threshold is not met then nothing gets notified (even if there is a match). Arguably this is why this approach is better for privacy. Cloud matches would not be able to have this extra threshold (in addition to this model allowing e2ee on the cloud in the future). I'd also like to know more about the specifics here, my guess is that threshold value is pretty high (their 'one in a trillion' comment not withstanding). It's probably targeting large CSAM dumps of matches which would not get flagged by different images.
slownews45 5y ago
Absolute - I think this is one of two key questions for me. That is why I put "known" in quotes. It can't be an exact match because it has to handle cropping, rotation, resize etc. Images then do get a manual review before a report is made which is good and may help provide feedback on alogs being used. Going to be hard though for apple to set the second factor to high - I'd say 5 maybe? It's hard to say you had matches on potential CASM and ignored them I'd think.
still_grokking 5y ago
According to https://rentafounder.com/the-problem-with-perceptual-hashes/ https://rentafounder.com/the-problem-with-perceptual-hashes/ the false-positive rate will be likely high. Given the billions of pictures going through this system there are going to be a lot of false accusations of child porn possession likely (and alone such an accusation can ruin lives). HN discussion of that article from a few days ago: https://news.ycombinator.com/item?id=28091750 https://news.ycombinator.com/item?id=28091750
- slownews45 5y ago
  This is where the thresholding and manual review come in, but could be a bit scary for sure.
  - cwkoss 5y ago
    Even with thresholding and manual review, the idea that most iphone users (who don't have any CP) are going to have a non-zero child porn score in apple's database is just super creepy. Just a minor policy change (lower threshold) away from calling millions of people pedophiles.
    - simondotau 5y ago
      The threshold is cryptographically enforced. See page 4 of Apple's technical summary.

by cwkoss 5y ago

fossuser 5y ago

It's two factors, both the match on an image hash and an unknown threshold of matches at which point the data gets sent up. If the threshold is not met then nothing gets notified (even if there is a match). Arguably this is why this approach is better for privacy. Cloud matches would not be able to have this extra threshold (in addition to this model allowing e2ee on the cloud in the future). I'd also like to know more about the specifics here, my guess is that threshold value is pretty high (their 'one in a trillion' comment not withstanding). It's probably targeting large CSAM dumps of matches which would not get flagged by different images.

slownews45 5y ago

Absolute - I think this is one of two key questions for me. That is why I put "known" in quotes. It can't be an exact match because it has to handle cropping, rotation, resize etc. Images then do get a manual review before a report is made which is good and may help provide feedback on alogs being used. Going to be hard though for apple to set the second factor to high - I'd say 5 maybe? It's hard to say you had matches on potential CASM and ignored them I'd think.

still_grokking 5y ago

According to https://rentafounder.com/the-problem-with-perceptual-hashes/ https://rentafounder.com/the-problem-with-perceptual-hashes/ the false-positive rate will be likely high. Given the billions of pictures going through this system there are going to be a lot of false accusations of child porn possession likely (and alone such an accusation can ruin lives). HN discussion of that article from a few days ago: https://news.ycombinator.com/item?id=28091750 https://news.ycombinator.com/item?id=28091750

slownews45 5y ago
This is where the thresholding and manual review come in, but could be a bit scary for sure.
- cwkoss 5y ago
  Even with thresholding and manual review, the idea that most iphone users (who don't have any CP) are going to have a non-zero child porn score in apple's database is just super creepy. Just a minor policy change (lower threshold) away from calling millions of people pedophiles.
  - simondotau 5y ago
    The threshold is cryptographically enforced. See page 4 of Apple's technical summary.