Efficacy of pixel swap-based steganographic algorithms in grayscale images

Abstract

The field of steganography deals with the encoding of hidden messages into other data, called the cover, in a manner that makes it non-obvious that a hidden message exists. Though steganography does not necessarily, on its own, encrypt the hidden data beyond deciphering, it can be used as a supplement to encryption, avoiding unnecessary attention from adversaries who may otherwise take additional measures if aware of the secret message. Many types of cover media may be used, ranging from text and formatted text documents, to audio and video, and images. Our study focuses on grayscale images as the choice of cover media.

Many kinds of image steganography already exist. The most common, Least Significant Bit (LSB) steganography, efficiently hides data within pixels’ intensity values, in a manner unnoticeable to the naked eye. Because LSB alters the cover image’s first-order statistics, however, it is often detectable through steganalysis methods such as Sample Pair Analysis. While less-detectable variants of LSB, as well as other methods, have been separately introduced, we focus on permutationbased methods that avoid this disadvantage through not alternating the first-order statistics to begin with.

Several pixel-swapping algorithms have already been introduced in the literature. We generalize upon those methods by allowing general permutations within larger sets of pixels and intensities, called permissible sets. By design, these permissible sets are an invariant, ensuring that both encoder and decoder read the same ones, even post-permutation. To serve as support for this technique, a supporting theory of multiset permutation is devised and applied.

The wider range of possible permutations increases the bit-per-pixel embedding rate over swap-based methods, in a manner that also reduces detectability. Direct implementation and comparison shows our method to improve upon previous swap-based steganography for the Microsoft Research Cambridge dataset of general images, for fixed bit-per-pixel rates. It also shows a larger improvement for the NoisyOffice dataset of scanned images.