JPEG Steganalysis Based on DenseNet

Current research has indicated that convolution neural networks (CNNs) can work well for steganalysis in the spatial domain. There are, however, fewer works based on CNN in the JPEG domain. In this paper, we have proposed a 32layer CNN architecture that is based on Dense Convolutional Network (DenseNet) for JPEG steganalysis. The proposed CNN architecture can reuse features by concatenating features from all of the previous layers with the same feature-map size. The shared features and bottleneck layers make the CNN improve the feature propagation and reduce the parameters. Experimental results on the BOSSbase, BOWS2 and ImageNet datasets have showed that the proposed CNN architecture can improve the performance and enhance the robustness. To further boost the detection accuracy, an ensemble architecture called as CNN-SCAGFR is proposed. Our method is the first time to integrate the CNN and conventional method in the JPEG domain. Compared with the current method XuNet [1] on BOSSbase, the proposed CNN-SCA-GFR architecture can reduce detection error rate by more than 4% while the number of parameters used is only 17% of what used by XuNet.