Statistics of Low-Order Bits in Images

[kqb at whscad1.att.com]

Several people are attempting to create an algorithm to mask the
presence of a steganized encrypted message in the least significant
bits of an image.  Don't forget that no matter how fancy your algorithm
or how closely you mask your steganography with a model of what the
statistics of an ordinary image look like, you have to assume that your
opponent also knows your steganization algorithm, including your masking
technique.  (Otherwise you are just relying on security through obscurity.)
This leaves you with three problems:

  (1) your opponent may have a much better model of an ordinary
      image than you do, and still be able to discern the existence
      of masked steganography,
  (2) since your opponent knows your steganization algorithm,
      he/she can look for any "signature" that your steganography
      masking model leaves, and
  (3) your opponent can "desteganize" all your images and check their
      statistics for deviations from the statistics for "desteganized"
      ordinary images.

Resolving problems (1) and (2) requires a lot of work constructing
good models.  Resolving problem (3) requires, I think, a modeling
function for steganography that is invertible only with a secret key.
(Otherwise, your opponent could desteganize your image and find a
uniform random distribution, which indicates an encrypted message.)
Since this type of function is, to my knowledge, not well-developed,
don't expect it to be secure.  Thus, if breaking it could compromise your
secret key for desteganization, then don't use the same public/private
key pair for both encryption and steganography.

                              Kevin Q. Brown
                              INTERNET    kqb at whscad1.att.com
                                 or       kevin_q_brown at att.com