An Advanced JPEG Steganalysis Method with Balanced Depth and Width Based on Fractal Residual Network

Digital image steganalysis technology based on deep learning has made rapid development. The latest technology, SFNet based on fractal technology, exceeds the cutting-edge technology SRNet in performance. The disadvantage is that SFNet is not suitable for JPEG steganalysis. Due to the interference of compression noise in JPEG images, it is difficult for SFNet to extract weak stego signal by self-similarity extended network. In this paper, a JPEG steganalysis method based on fractal residual network, called FRNet, is proposed. This paper introduces the residual unit with a shortcut connection into the fractal structure, which helps the network effectively suppress the image content and generate the residual image with stego noise. Then, referring to the bottleneck block of ResNet, the deep feature extraction module is constructed to downsample the feature map and superimpose the weak stego signal between different channels of the convolution layer. Finally, the fractal residual module and depth feature extraction module are used to control the width and depth of the network to maximize the detection performance. Two adaptive steganography algorithms of J-UNIWARD and UERD are chosen to evaluate the performance. Experimental results show that the detection error of FRNet is 11.52% lower than J-XuNet, 10.12% lower than WangNet, and 2.54% lower than SRNet.