JPEG steganalysis using color correlation and training on clean images only

Steganalysis has becoming an emerging important technique for detecting secret messages that are embedded in a clean-image. Universal steganalysis is especially useful due to its independence of prior knowledge of the embedding procedure. However, the detection results from the majority of universal methods are largely determined by the training procedure on a mixture of clean-images and stego-images, and therefore not practically feasible. Moreover, many color steganalysis methods do not take color coefficients into special consideration and thus they can be viewed as a simple extension of the analysis for grayscale images. To capture the distinct features of the clean images, we propose a novel predictor based on the intra- and inter- color correlations of wavelet coefficients. This method achieves higher detection rates, under a blind condition that only involves clean images at the training stage.