> If someone were to release code that reverses NCMEC hashes into pictures, then everyone in possession of NCMEC's PhotoDNA hashes would be in possession of child pornography.
Please correct me if I'm wrong, but wouldn't it be more correct to say they "would be in possession of images recognized by PhotoDNA as child pornography" rather than actual CP?
Technically not possible without severely grasping assumptions. It would be more likely you would create a collision and have calculated an image of a duck (south african whitenoise duck).
Problem with all this is that the images of naked children made by their parents is CP in the eyes of its consumers.
Perceptual AI is the best approach, but produces a certainty < 1.
In my cryptography course I had a project about invisible watermarks and secret messages in imaging. The first hurdle was beating partial images and compression, so most early algorithms worked in the frequency domain. At that time it was basically an arms race between protection or deletion of said messages and I think that hasn't changed.
Conventional file hashes can be beaten by randomizing meta data since a quality hash function would immediatly create a completely different hash. Never mind just flipping or probably just resaving them.
If you create a polynomial approximation of frequencies or color histograms of an image, you have a relatively short key indicator. But you need a lot of those to even approach certainty. Could always be an image of a duck.
Please correct me if I'm wrong, but wouldn't it be more correct to say they "would be in possession of images recognized by PhotoDNA as child pornography" rather than actual CP?