Abstract
Recently steganalysis methods based on convolutional neural networks (CNN) have achieved great improvement. However, detection against adaptive steganographic algorithms with low embedding rates has still been a challenging task. To deal with this problem, we propose a CNN steganalysis model employing the joint domain detection mechanism and nonlinear detection mechanism. For the joint domain detection mechanism, we use not only the high-pass filters from the SRM for spatial residuals, but also the patterns from the DCTR for frequency steganographic impacts. For the nonlinear detection mechanism, we enlarge steganographic effects by nonlinearly transforming the extracted steganographic residual information. In addition, we innovatively put forward a model learning method based on the high learning ability of a model. That is, we use lower embedding rate image datasets to train a model and higher embedding rate image datasets to test the model, which effectively improves sensitivity to steganographic traces. Compared with the existing steganalysis models such as SRM+EC, Ye-Net, Xu-Net, Yedroudj-Net and Zhu-Net, the detection accuracy of our model is about 4%∼6% higher than that of the best Zhu-Net model.