For imagesteganalysis, many studies have showed that the superiority of the convolutional neural network overconventional methods based on artificially designed features. Withthe trend of the fusion of traditional steganalysis methodsand some tricks used in classic computer vision tasks, such asSRNet equipped with residual modules and ZhuNet which usedspatial pyramid pooling, more and more CNN architecturesused for steganalysis are proposed. However, there still are somecharacteristics in most content-adaptive steganographic algorithms such as S-UNIWARD, HUGO, WOW, and tricks in designing network structure whichcan be used for steganalysis. Here, we propose a CNN network framework which can further improve theperformance of spatial imagesteganographic algorithms. First, we utilizemore SRM kernels to initialize the pre-processing layer than previous CNNs, and usean image padding method different from traditional modelsto preserve the integrity of image residuals as much as possible. Next, we use multiple channel attention layers which aim to discriminate the more informational features boosting the detection accuracy of network. Then, we deploy the spatial pyramid poolinglayer before features are fed into the fully-connected layers, aiming to extract more features from the last feature mapsin several scales. Several experiments under different steganographic algorithms show that, the proposed CNN outperforms the other CNN-based steganalyzerssuch as YeNet, XuNet, YedroudjNet,SRNet and ZhuNet.
A Novel Spatial Image Steganalyzer with Adaptive Channel Attention
