Hi all<p>As I scanned the daily news with more amazing image generations from the new kid on the block, and the usual comments and posts that fear doom and gloom over the AI future(fake images, fake videos, etc), I couldn't help but wonder why we still don't use reliable means to provide verification of an image authenticity.<p>A quick google search revealed papers as old as 1998[1], where a proposal for using public keys for that purpose was presented.<p>Picture this, you see some shocking image of Elon Musk having a date with a humanoid robot (I actually saw this photo yesterday). But now you have a tool to submit that photo for verification. Maybe sources like Getty, AP, CNN, etc, would have their public keys available for anyone to cross-check the authenticity of images. Much like we do today with PGP/GPG.<p>Perhaps a whole new image format could be developed that would even facilitate this(or require such keys to be used). And there would be no gatekeeping, everyone can have their own private/pub keys, like we already do now. Famous photographers will have their pub keys on their websites so that people can use them to verify.<p>If AI-generated images are such a problem (and will become a bigger one), why is this not being done?<p>[1] - https://ieeexplore.ieee.org/document/723526
There are lots of problems with this approach, but one of the bigger ones is that if everyone has their own keys, then everyone can create signed images that have been manipulated.<p>So even if you you can get people to actually verify images (and bear in mind how hard it is to get people to manage keys and use something like GPG), not to mention the technical issues with image reproduction, this would only give you protection in cases where someone is claiming that an image was created by a specific individual and it wasn't. But that's not the problem in the <i>vast</i> majority of instances.
You're not testing what you think you're testing. We can verify an image, but still have problems with the story below the image. I don't think this should be half-solved, it would give false credence to real images used in lies. Or the reverse, you discredit a fake image of Musk you saw yesterday, but tomorrow he does it for real.<p>You'd need to give political campaigns' keys some trust, for images you expect from them. But what if they start signing images that you wouldn't expect from them, then you have 'verified' fake news.
There’s actually a web standard to authenticate images.<p>Several cameras on the market implement it, and even openai authenticates their images.<p>The limit of signing an image is that any modification will break authentication, but zero knowledge proofs can be used to “preserve authentication”. The current signatures & zkp have some practical limitations for large images, but nothing that won’t be fixed soon, if I had to guess.<p>I recommend this presentation: <a href="https://youtu.be/EKoY8ysGblk?si=3-lLrzCP7263sY_J" rel="nofollow">https://youtu.be/EKoY8ysGblk?si=3-lLrzCP7263sY_J</a>
What is the problem solved or revenue opportunity from image verification and authentication?<p>A fake image of Elon doing whatever would general billions of clicks, ad impressions, etc., which is all that matters.<p>If you want to get only genuine images from AP, CNN, etc., just go to those websites (using HTTPS).