TCNN: Two-way convolutional neural network for image steganalysis

Recently, convolutional neural network (CNN) based methods have achieved significantly better performance compared to conventional methods based on hand-crafted features for image steganalysis. However, as far as we know, existing CNN based methods extract features either with constrained (even fixed), or random (i.e., randomly initialized) convolutional kernels, and this leads to limitations as follows. First, it is unlikely to obtain optimal results for exclusive use of constrained kernels due to the constraints. Second, it becomes difficult to get optimal when using merely random kernels because of the large parameter space to learn. In this paper, to overcome these limitations, we propose a two-way convolutional neural network (TCNN) for image steganalysis, by combining both constrained and random convolutional kernels, and designing respective sub-networks. Intuitively, by complementing one another, the combination of these two kinds of kernels can enrich features extracted, ease network convergence, and thus provide better results. Experimental results show that the proposed TCNN steganalyzer is superior to the state-of-the-art CNN-based and hand-crafted features-based methods, at different payloads.