Okay, there are watermarks on AI stuff, interesting, not good but everyone else already talked about.
My question..how does it work? Is it robust? I remember that young me hid data in pixels of png with simple stegonagraphy and it was a fun little project..but brittle. How exactly does the fingerprinting survive jpeg compression? Is it repeate over and over the images or is it just one area? If that one is pure black by chance the jpeg algorhitmen would erase it all no?
I'm pretty sure the robustness of watermarking has been a solved problem for at least 30 years, I remember visiting the mit media lab in the mid 90s and they were explaining how they could watermark digital audio by imperceptibly adjusting the acoustics as if the walls of the room the audio has been recorded in were changing distance relative to the microphone according to a wave function
There is no robust-without-qualifiers watermarking. There is only robustness against common operations but as soon as someone is actually trying to remove the watermark they can at the very least use exactly the same technique to embed random noise instead to overwrite your watermark.
My question..how does it work? Is it robust? I remember that young me hid data in pixels of png with simple stegonagraphy and it was a fun little project..but brittle. How exactly does the fingerprinting survive jpeg compression? Is it repeate over and over the images or is it just one area? If that one is pure black by chance the jpeg algorhitmen would erase it all no?