Dual Convolutional Neural Network for Image Steganalysis

Abstract

In this paper, we propose a new steganalytic method that uses dual convolutional neural network (CNN) of which each has different inputs. To construct the dual CNN structure, two pairs of the preprocessing filters and the convolutional layers were brought from the conventional CNN-based steganalytic methods and the outputs of the dual CNN were concatenated and fed together into a following affine layer. Given an input image, a stego image is created by embedding some additional data into the input image using one of steganographic methods and a difference image is computed between the input and stego images. Then, the input and difference images are fed into each CNN, respectively. This indicates that the proposed method extracts /learns additional features from the difference image using the additional CNN. Experimental results demonstrated that the proposed dual CNN with additional input can identify whether the S-UNIWARD steganography was applied to the input image with an accuracy of 80.43%, and can improve the accuracy by approximately 5% when compared with the conventional CNN-based steganalytic method.