Model-based steganalytic method towards color JPEG images

JPEG images have become an important type of cover medium for steganography. Steganalysis of JPEG images has received much attention in recent years. But most of the known steganalytic algorithms are targeted for specific embedding techniques and steganalysis of color images are less covered. In this paper, we present a new universal steganalytic method of detecting the existence of secret message in color JPEG images. In YCbCr color space, two-dimensional block discrete cosine transform (DCT) is performed on the blocks of each color component of JPEG images. The DC and AC coefficients are modeled as Gaussian distribution (GD) and generalized Gaussian distribution (GGD), respectively. Relative entropy is used to measure the difference between the distributions of the DCT coefficients of cover images and those of stego-images. The changes in the distributions of the DCT coefficients caused by several popular DCT domain steganographic techniques are studied. The distributional parameters of the mixture statistical models of DCT coefficients are extracted as distinguished features to discriminate cover images and stego-images through a support vector machine (SVM) classifier. Experimental results show that the proposed steganalytic method can attack color JPEG image steganography successfully with high accuracy. And the classifier based on the proposed distinguished features behaves more efficient than that based on the features proposed by Farid.