3 ms·

I was thinking about this the other day. The issue is if you have an imperfect test to detect fakes, it gives even more credibility to fakes that pass the test.

I was thinking about this the other day. The issue is if you have an imperfect test to detect fakes, it gives even more credibility to fakes that pass the test. If there are no tests however, then were left to question the validity of everything

michaelbuckbee 3y ago
What makes it even more of an issue is that's comparatively easy to generate a 1000 images of a scene and push them all through until you get one that happens to line up in such a way as to pass through the detection (as compared to having to physically paint 1000 scenes).
- andoando 3y ago
  Right exactly, much like p hacking with studies. Publish enough research papers and you get significant results by chance, and it gets into a headline as "Science shows..." and people believe it. Unfortunately in this space the amount of deepfakes we can create is massive, so even a system thats 99.999% accurate will leave a lot of fakes through and grant them credibility. I think we have to focus on an alternative path where we assume any digital content can be fake, unless its creation is provable through some verified sources or methods. Maybe it's possible to embed some kind of digital fingerprint into images? So as to say, this image was definitely taken by this camera. Pretty sure something like this has been done.
  - michaelbuckbee 3y ago
    I'm not even sure the "digital fingerprint" is entirely helpful. Have you seen the moon issue? https://www.reddit.com/r/Android/comments/11nzrb0/samsung_space_zoom_moon_shots_are_fake_and_here/ https://www.reddit.com/r/Android/comments/11nzrb0/samsung_sp...
    - andoando 3y ago
      I ended up reading that whole thing, but by digital fingerprint I meant something like embedding a cryptographic checksum in an image. Or some way of proving an image is produced by a specific camera. Itd be great for example if we could verify a security image video was real and not a fake
8A51C 3y ago
Continuing this thought, consider a photo may be genuine but the actual scene is faked. Will pass all tests.

michaelbuckbee 3y ago

What makes it even more of an issue is that's comparatively easy to generate a 1000 images of a scene and push them all through until you get one that happens to line up in such a way as to pass through the detection (as compared to having to physically paint 1000 scenes).

andoando 3y ago
Right exactly, much like p hacking with studies. Publish enough research papers and you get significant results by chance, and it gets into a headline as "Science shows..." and people believe it. Unfortunately in this space the amount of deepfakes we can create is massive, so even a system thats 99.999% accurate will leave a lot of fakes through and grant them credibility. I think we have to focus on an alternative path where we assume any digital content can be fake, unless its creation is provable through some verified sources or methods. Maybe it's possible to embed some kind of digital fingerprint into images? So as to say, this image was definitely taken by this camera. Pretty sure something like this has been done.
- michaelbuckbee 3y ago
  I'm not even sure the "digital fingerprint" is entirely helpful. Have you seen the moon issue? https://www.reddit.com/r/Android/comments/11nzrb0/samsung_space_zoom_moon_shots_are_fake_and_here/ https://www.reddit.com/r/Android/comments/11nzrb0/samsung_sp...
  - andoando 3y ago
    I ended up reading that whole thing, but by digital fingerprint I meant something like embedding a cryptographic checksum in an image. Or some way of proving an image is produced by a specific camera. Itd be great for example if we could verify a security image video was real and not a fake

8A51C 3y ago

Continuing this thought, consider a photo may be genuine but the actual scene is faked. Will pass all tests.