Hacker News Reader

Comment by topaztee | Hacker News Reader

Is there some hidden signature behind this AI generated content so we can differentiate whats real and what isnt?

Short answer, no, there's not. Adversarial AI could be used to find deepfakes, but then AI will be trained on those adversaries etc.

There is no real solution to this problem, but my thoughts on the matter are that in the future it will be common for cameras or recording devices to have a cryptographic signing key and the resulting footage would contain a signature. This would allow news outlets to prove to their audiences that the footage is untampered with. It could also be chained so that officially licensed syndicates could also offer provenance to their audiences, or that clips/re-encodes of the same footage could link back to an original recording.

This doesn't stop organisations or individuals from recording a deepfake, but it would put their reputation on the line (similar to how we have certificate authorities signing SSL certificates now).

It's up to you then if you want to trust the footage provided by AP vs. the footage provided by Infowars.

Like the new Leica M11-P with its "content credential"?

Yes, very much along the same lines. Though proving that a photo was taken with a certain camera is not really enough. You need to prove that it came from a particular photographer/journalists camera who licensed their images to an outlet who is willing to certify that they trust the photographer not to have tampered with the image (or along this type of line).

Certainly you could also just trust the photographer, but a chain of trust might be preferable to the general viewer (eg. I don't necessarily trust Photographer X, but I trust The Guardian and if The Guardian trusts this photo taken by Photographer X to be authentic, then I trust this image to be authentic).

Edit: I read more about the CAI and C2PA[1] and I think it would be sufficient for provenance and origin purposes, so what is missing is the trust network. It's likely going to a while for media outlets to establish trust networks and phase out the old, untrusted ways. There's definitely going to be room for both and I imagine outlets will start using terms like "we have received images from multiple sources, however their authenticity has not be verified by our editors".

[1] https://c2pa.org/specifications/specifications/1.3/specs/C2P...

Currently there are tiny imperfectios but I believe in 10 or 20 years they will be indestinguishable with the naked eye.

Just eyeballing it for a minute works for me. But people are prejudiced and their opinions will always cloud their judgement, no matter what the evidence or counter-evidence is. So overall, nothing will ever change.

I feel that the being able to tell by studying it for a bit is only going to get harder and harder as the technology advances

Reply on news.ycombinator.com