[dimas_codes]

I am sorry if I missed something or someone already asked it, but:

If I generate image with AI, print it, then take a photo of it with Roc Camera so that you can't tell that this is actually a printed image, I will then have an AI image with ZKP of its authenticity?

[dbdr]

Sony has this on their related page:

> A digital signature alone cannot determine whether the captured image is of an actual 3D subject, or of an image or video projected on a high-definition monitor. However, by using metadata including 3D depth information, it is possible to verify the authenticity of images with a high degree of accuracy. By using cameras from Sony, both the image and the 3D depth information can be captured on the sensor along the single light axis, providing information of high authenticity.

That 3D depth data could presumably be used to detect this. In principle, you could also train an AI to generate realistic 3D data. It's just not available yet, and probably harder to train (in general, and also since you would need to collect new massive amount of training data first).

No idea if this specific device has a 3D sensor, addressing the general question.

[zipy124]

The depth information that sony cameras collect is almost certainly low-res enough that even with a simple image->depthmap model[0] you could fool it. Also they don't say anything about the sensor itself being secure, no need to print something if you can just emulate the sensor with an FPGA or other.

[0]: https://github.com/CompVis/depth-fm

[efskap]

I suspect the EXIF data won't make sense, and the faq says the ZKP applies to the metadata as well. But yeah, inherent flaw.

[keiferski]

Presumably you could stop this by requiring GPS data for the image, and match that against a library of other images in the location?