CNN-based image steganalysis using additional data embedding

Abstract

Image steganalysis identifies whether a secret message is hidden in an image. Conventional steganalytic methods require processes to extract discriminative statistical features from images and classify them. Convolutional neural networks (CNN) are particularly effective at conducting those processes. However, since the hidden message was too weak to be detected, existing CNN-based steganalytic methods needed to adopt preprocessing filters to increase the strength of the hidden message. Then, development focused on improved network structures and preprocessing filters. In this paper, we propose a different approach to CNN-based image steganalysis. We embed additional data in an input image and use two images (i.e., the original input image and its stego image with additional embedded data) as input. This is based on an assumption that pixel variations due to the additional embedded data would be sufficient to identify images with and without a secret message. We also propose two variants of conventional CNNs for image steganalysis, named dual channel CNN and dual network CNN, to input two images. We conducted various experiments using the proposed CNNs. The experimental results prove that the assumption holds, and the additional input could provide useful information to improve the performance of conventional CNN-based steganalytic methods. Depending on the strength of the hidden message, the proposed approach could improve the identification rate by up to 6% for S-UNIWARD, an adaptive steganographic method.