Steganalysis Using Unsupervised End-to-End CNN Fused with Residual Image

Abstract

Recently, convolutional neural networks (CNNs)has been used in the field of image steganalysis. However, there are still many deficiencies. In order to improve the detection accuracy, we propose an unsupervised end-to-end CNN to extract image features of the stego images. The end-to-end mapping can be trained to learn the most effective characteristic expression from input images to output images. By integrating hidden layers of the deep CNN, the extracted features can be considered as having characteristics of both input images and its residual images. In this way, we try to minimize the negative effect of the high-pass filtering under the condition of guaranteeing the convergence of the network. The experimental results show that the end-to-end CNN maintains good performance on BOSSBase even when the embedding rate is 0.1 bpp.