Research On Image Steganalysis With Low Embedding Rate Based On Convolutional Neural Network

The idea of image steganography is to embed secret information in redundant areas of the image so that humans cannot perceive changes in the image.The main purpose of the image steganalysis is to detect whether the image is hidden from secret information.With the development of adaptive spatial steganography algorithm,the limitations of feature design of traditional steganalysis methods are becoming more and more prominent.As a new steganalysis solution,convolutional neural network has become a research hotspot in the field of steganalysis.In recent years,the research of spatial steganalysis based on deep learning has achieved good results,but the detection performance under low embedding rate is still not ideal.Low embedding rate means little secret information is hidden.Therefore,this paper focuses on the problem that the adaptive spatial steganography algorithm is difficult to detect correctly under low embedding rate,studies and proposes two steganalysis methods as follows.(1)A new convolutional neural network structure is proposed for low-embedded rate image steganalysis.This network uses 30 SRM filters for preprocessing to obtain implicit noise residues,adopts three convolution layers and reasonably designs the size of convolution kernel to extract weak steganographic feature under low embedding rate.The performance of the network is improved by appropriately selecting the batch normalization operation and the activation function,and the loss of the noise feature is avoided by canceling the pooling layer.Based on the network structure,a transfer learning method called step-by-step is proposed.By gradually migrating the model parameters under high embedding rate to the low embedding rate training task,the detection effect under low embedding rate is further improved.The experimental results show that compared with the existing methods,the proposed network can achieve better detection results,especially under the low embedding rate of 0.05 bpp.(2)A network structure based on channel selection is designed to improve the detection effect of image steganalysis under low embedding rate.Based on the convolutional neural network designed in this paper,a channel selection module is introduced,which can extract more prior information from the input image by using the known embedding algorithm and embedding strength,and enhance the steganalysis analysis capabilities of the model.In the channel selection module,the embedded probability map of the input image is calculated first,and then the weight map is extracted from the probability map by using the SRM high-pass filter kernel.Finally,the weight map is used to weight the noise residue image outputted from the preprocessing layer.By assigning different weights to different pixels of the residual image,the network can extract features from the texture and noise residual regions more effectively.The experimental results show that the network with channel selection module has better detection capability.