In the last post I demonstrated how a the origin of an image could be concealed by small changed to the quantization tables. In this article I'm attempting to make a drawing look like a photograph - at least in regards to its quantization tables.
I will start from the following drawing:
dqtmd5: 620eb457891670b5b9f9504d10d32bf4
Now I'd like to change the quantization tables to resemble those that are typical for a Canon EOS 10D. For this I open the drawing in PhotoFiltre and save it at a custom quality level of 97%. At this quality level the quantization tables come close to those of a 10D. Next I copy and paste the tables from a reference image into the drawing. And here is the result:
dqtmd5: d2c2c9f7b759bd6ffa62f57a28381ac4
Of course the dqtmd5 sum now indicates a 10D. The smiley face looks pretty good, but lots of artifacts clearly can be seen around the letters. So in this case it's easy to uncover the forgery.
So a thorough examination of the visual impression would be inevitable to detect manipulations. How can this step be done without user intervention? In this case edge detection could have been helpful. But does this apply to all sorts of quantization table manipulations? Further research is needed...
