[mcintyre1994]

What hash algorithm are they using? If this is legitimate (I’ll be honest - I’m not clicking) then surely any hash this easy to pre-image attack is completely useless? Why wouldn’t they be using a cryptographic hash here?

[soneil]

As I understand it the NN "perceptual hash" is supposed to hash the image, not the file. eg, if I take a photo of my cat, and hash the file. Then remove my geo data from the exif - the hash no longer matches. It is still very clearly my cat, but cryptographic hashes don't match. This could be resizing the image, saving as png, mirroring/flipping it, etc.

The "perceptual hash" should be able to say "no, that's still the same image" while the file data has been entirely transformed.