Maximizing steganalysis performance using siamese networks for image

Image steganalysis is used to detect the presence of hidden data. Recent studies have shown that deep convolutional neural networks (CNNs) applied to steganalysis exhibit excellent performance. However, current network architectures have deepened layers to pursue an ultimate local receptive field, overlooking the boundary and overall information of the image. As a result, the network fails to effectively extract steganographic feature information. In this paper, we propose a method that effectively captures both boundary and global information. We process the images through segmentation and padding, followed by treatment with four symmetric sub-networks with shared parameters and structures to acquire more comprehensive steganographic features. By integrating two loss functions into the traditional cross-entropy loss, we can train a more compact feature space, thereby enhancing network performance. Experiments were conducted on the BOSSbase1.01 dataset, using two widely employed steganography methods, namely WOW (wavelet obtained weights) and SUNIWARD (spatial universal wavelet relative distortion), for comparison. Results show the proposed model demonstrates superior performance on various payloads.