This is exactly a technique used in media forensics. You can even tell things like crop region and lens geometry (due to subpixel statistics and chromatic aberr

This is exactly a technique used in media forensics. You can even tell things like crop region and lens geometry (due to subpixel statistics and chromatic aberration). Train a resnet on tens/hundreds of thousands of images and you'd be surprised what labels it can learn.