If you're interested in ELA, I'd encourage you to try it on some control photographs you've taken yourself, and know are unmodified, particularly photographs that have objects with considerably different textures, like textured fabrics.
Trying it on lossless screenshots can also be amusing.
ELA does work when someone manipulates an existing JPEG. If you take a low quality JPEG, paste in part of a high quality JPEG and save in high quality you can easily tell which area has been modified.
Performing ELA on the output JPEGs when the editing has been performed on the high resolution raw images in Photoshop does not work. The down-sampling process effectively normalizes the noise levels.
The ELA analysis in this article is pure nonsense, it is simply highlighting the areas with high frequency components such as the dress texture, fine hairs and features on the face.
To do a proper study like this you need to compare against a sample of similar images which you know have not been manipulated.
In this case, isn't there a sort of "end justifies the means" applicable to this analysis, coming from the fact that the author now has access to the (closer to) original photo?
If you're interested in ELA, I'd encourage you to try it on some control photographs you've taken yourself, and know are unmodified, particularly photographs that have objects with considerably different textures, like textured fabrics.
Trying it on lossless screenshots can also be amusing.