I was always fascinated by stuff like LSB stenography where you hide a message (or another image) in the least significant bits of an image. Not visible to the human eye, mathematically should look like thermal noise in the image at most.

Image sensor noise is not a uniform distribution (I believe it's typically modeled as a Gaussian distribution, but even that's an approximation).

Thus, it is mathematically/statistically distinguishable.

The LSB of samples from a gaussian distribution is uniformly distributed: 50/50 chance of 0/1

(Assuming you have enough bits per sample which image sensors provide IRL)

Only in theory. In practice, they're correlated with the other bits and adjacent pixels.

How correlated? I spent 2 minutes trying and failing to get chatGPT to generate code that downloads an image dataset and runs diehard on the LSBs. I'd expect the LSBs to pass the randomness test.