No, it doesn't work like that.

zionic · on Aug 12, 2021

Yes it does, it uses fuzzy perceptual hashes not crypto hashes.

So if your innocent baby pic looks similar enough to a previously tagged child abuse image then YES, it will flag you and send a copy to the feds.

And before you correct me, the Apple employee will see a picture of your naked baby and hit “forward to NCMEC”, which… upon investigation is actually just the feds

nicce · on Aug 12, 2021

Perceptual hashes are extremly accurate. You might need a twin and top of that somehow identical environment almost in pixel level.

Are news filled with false-positive accusations by PhotoDNA, flagging wrong images in Google, Facebook, Instagram etc.?

heavyset_go · on Aug 13, 2021

> Perceptual hashes are extremly accurate.

No, they are not. Ask anyone who has worked in this space[1][2], including myself. They are incredibly common.

Two images that kind of look like one another will have similar or the same hashes. That is the point of perceptual hashing.

[1] https://hackernews.hn/item?id=28091750

[2] https://hackernews.hn/item?id=28110159

outworlder · on Aug 12, 2021

IF there are multiple matches, IF it's going to icloud, THEN a 'derivative image' will be show for screening and IF deemed to be warranted, sent to NCMEC.