Something else you can do is use a cipher which takes two input streams and merges them into the one file, with one key extracting the 'harmless' information and another extracting the 'harmfull' information.
AFAIK, the only way to do this is with a Vernam OTP. You have a key file (A) the same length as your real data (B) -- encrypt the data by XOR to get (C). Then you take an innocent text (D) and XOR with (C) to get an alleged key file (E). You hide (A) someplace, destroy (B) and (D). Leave (C) around and put up just enough resistence in letting folks have (E).
Does anyone know a simpler way? I'm willing to bet that it can be proved that the key would have to be at least 1/8 the length of the message in order for this to work but I don't know of any schemes using less than the message length to do it.
Yes. Make "noise addition" (really multiplexing) part of the cipher. You could throw away every other bit based on the parity of the key. The ciphertext would be twice as big, but if you compressed both plain text streams first, this effect might not be very obvious. Of course, if your encryption program were disassembled, you might be found out...