Parametric Approaches to Steganalysis, 16-R9632Printer Friendly Version
Inclusive Dates: 05/06/06 05/15/07
Background - Steganalysis is the art of finding information intentionally hidden in an otherwise innocuous medium through means of steganography. The recent explosion of digital communication, in particular the Internet, has created an enormous capability for distributing multimedia content. In this world, covert communication through steganography becomes a simple matter. Perfectly reconstructable digital documents, powerful multimedia processing tools, and worldwide access to near instantaneous, near anonymous communication has created an ideal medium for covert communication through the subtle and near undetectable techniques of digital steganography.
Approach - Just as with encryption, the battle between steganography and steganalysis is never-ending. This is a leading-edge research effort to explore the field of steganalysis. In this effort, two fundamental approaches are under examination. The first is a targeted effort to locate an embedded message by assuming an embedding method and training a noise removal system to detect this specific method. The second is a blind approach to developing a model of a non-embedded image, then using this model to detect deviations from the normal image. In both cases, the approach taken is not media specific, rather it is based on modeling the image itself or the "stego-information" embedded in the image.
Accomplishments - A series of experiments has been conducted using targeted noise removal techniques. Individually little success has been achieved over previously published results. The level of change using modern steganography techniques is simply not sufficient to effectively train a modern noise removal technique. The level of noise introduced through steganography is much too low. However, using a series of statistical tests in combination with these noise removal, or more accurately, clean image estimation techniques, has proven to be an effective technique for detection. This has led to a study of what makes an effective clean image estimation technique for blind steganalysis, as well as a fundamentally new approach to steganalysis based on modeling the very small amount of noise present in the natural image.