My understanding is that you can generate a statistical model of what a "natural" image is supposed to look like, and if something deviates from this model, it

My understanding is that you can generate a statistical model of what a "natural" image is supposed to look like, and if something deviates from this model, it may tell that there's something hidden in it. It doesn't say what is hidden or how to decode it, but it can indicate that something is there. E.g. there are certain coefficients in a JPEG that should fit a symmetrical curve, and the old JSteg tool used to break that symmetry while flipping bits to hide the payload. You couldn't tell by just looking at the image, but there were clear signs in the statistics.