To begin, please someone correct me if I am wrong, but I believe OutGuess was definitively broken in the paper "Attacking the OutGuess" in 2002 and this attack was never patched. It seems that since then, messages embedded with OutGuess have been detected reliably with other approaches as well.
Assuming the above holds, I thought it would be a fun project to make my own detector, starting with implementing the paper above. I have run into two issues:
- The paper describes and presents results for the grayscale case and mentions that it is "obvious" how to extend to RGB. It's not exactly obvious to me. Would I repeat the steps in the paper for each color channel (translated to YCbCr first I assume)?
- The paper seems to assume that everyone is careful when embedding messages and always uses an image large enough to allow for correction to defend against what broke OutGuess 0.1 (according to Wikipedia). What if this is not the case? Would the old attack still work, or is OutGuess able to provide some partial protection and what steps might I take to attack these types of images?
In all cases using an attack that allows me to estimate the message length would be really really cool. Or perhaps, are there already methods that allow extracting such information if you know with high probability that there IS a message?
