Reference Image Generation Algorithm for JPEG Image Steganalysis Based on Convolutional Neural Network

As the opponent of image steganography, the image steganalysis is to detect the secret message in images concealed by steganography algorithms. Recently, state-of-the-art JPEG image steganalysis schemes are changing from complex handcrafted feature-based ones to deep learning-based ones. Although the deep learning steganalysis for detecting JPEG steganography achieves great advancement, there still exists room for improvement. As it is verified that side information could promote the steganography detection accuracy, we seek the method to further improve the accuracy of content-adaptive steganography detection in JPEG domain from the perspective of side information offering for the deep learning steganalysis scheme. The proposed method utilizes convolutional neural networks to generate reference images from the input data. And the reference image is treated as the side information for the deep learning-based JPEG image steganalysis model. The proposed method can be pre-trained or trained together with the steganalysis model. Experimental results on classic content-adaptive steganography algorithms in JPEG domain named J-UNIWARD and JC-UED verifies the proposed method could enhance the detection ability compared with the deep learning steganalysis model without the aid of the proposed method to a certain extent. The proposed method could boost the detection accuracy for deep learning-based JPEG steganalysis model by 6 percentage points at most.