They probably produce "fingerprints" that are more resilient to small stenographic changes like a pixel change or even metadata changes or resizing the image, etc. They probably wouldn't fingerprint the data, but some representation of the data.
These always have tradeoffs, usually a false positive rate. Now, take someone like Facebook who has trillions of images (I dunno how many they have).
They should definitely be accountable, because I'm guessing the CP is only the form that FB can deal with (it's not public but in groups, messages etc)
They probably produce "fingerprints" that are more resilient to small stenographic changes like a pixel change or even metadata changes or resizing the image, etc. They probably wouldn't fingerprint the data, but some representation of the data.
The problem is that these are all usually stored in some database with this: https://en.wikipedia.org/wiki/Bloom_filter#Examples
These always have tradeoffs, usually a false positive rate. Now, take someone like Facebook who has trillions of images (I dunno how many they have).
They should definitely be accountable, because I'm guessing the CP is only the form that FB can deal with (it's not public but in groups, messages etc)