I think these tests are relatively easy to bypass, for instance by only
using certain pixels to hide data in and modifying the other pixels to
cancel out any bias.
A passage in Neal Stephenson's novel Cryptonomicon discusses this, and whole chapters are concerned with the interesting events around implementation of this - not with respect to a single image but to the real world.
If you've not read it I recommend it.
Ross Anderson's textbook Security Engineering is on the Newton College webserver somewhere under rja14 IIRC and is also worth reading and I think rather well-written. The second edition is in print, I have a paper copy of the first and may yet upgrade.
The first manoeuvre cryptologists apply to a message, I think, is to look for deviations from random. Arranging for there to be some which severely mislead them may be a game played by people who know more about it than I.
