Discussion
I am utterly intrigued by AI image provenance watermarking... It represents a new level of steganography that seems to take into account human psychology to throw people off with false positives and seemingly meaningful structures that don't correspond with anything on their own.
I wish I could meet the people who designed it and pick their brains.
Wait, so the goal is I can't tell it's AI, but someone with the secret knowledge could tell it's AI? That's awful!
@cy there's a verifier available but yes, you have to use their verifier and it just tells you yes/no, while the actual watermark contains a payload that tells the labs more. How much it contains is unknown but this is part of the spec for SynthID, if you want to look it up.
I'm designing my own provenance test to be able to tell if an image is AI generated, but I'm still early in the stages of research and don't want to spread and misinformation because I only have a slim grasp of the mechanics.
It's simple to include information outside of human visual range in an image. The problem is any provenance test you made, the people in control of AI could adjust to fool, and any provenance test they adjusted in the opposite way would work better. So only their cronies would be privileged enough to tell.
It's kind of like if only cylons can recognize other cylons.
@cy the SynthID marking *is* content aware, so in more complex images it gets more difficult to isolate, but I'm close to cracking that egg. Once I get it open I can keep diagnosing it like a black box until I have a better sense of how the encoder locates the watermark and what it may possibly contain. After that... it's cold hard math, which I would need to tap on someone smarter to determine things like distributed relationships of pixels across the entire image. It's not uncrackable, that's for certain. Especially if a fool like me can figure it out.
Install bonfire.cafe
Get the full app experience
