Steganalysis Based on Co-occurrence Matrix of Differential Image

A universal steganalysis method is proposed in this paper. The forward difference is calculated in three directions, horizontal, vertical and diagonal, towards adjacent pixels to obtain three-directional differential images for a natural image. Then the differential images are thresholded with a pre-set threshold to remove the redundant information. The co-occurrence matrixes of thresholded differential images are used as features for steganalysis. Support vector machines (SVM) with RBF kernel are applied as classifier to distinguish between stego images and cover images. The experimental results have proved that the proposed steganalysis method is effective in attacking the popular steganographic schemes applied in spatial domain. The detection rate for Cox´s spread spectrum (SS) data hiding method is above 78% over an image dataset consisting of 600 grayscale images. For generic LSB and plusmn1 methods, the proposed method achieves a detection rate above 95% as the data embedding rate is 0.3 bpp (bit per pixel).