Research On Adaptive Steganalysis Technology Based On Multi Channel Residual Learning For Color Images

Steganography algorithm is one of the common means for secret communication,but its abuse has brought a great threat to national and social security.The research of steganalysis technology is of great significance to timely discover and prevent the spread of digital content embedded with secret information.The adaptive steganography algorithm has high detection resistance and is the current mainstream.At present,there are many related studies on steganalysis for gray images,but relatively few related studies on steganalysis for color images.As one of the most widely disseminated digital content on social networks,color images are often used as the carriers for embedded information.Therefore,the research on adaptive steganalysis technology for color images is of great help to the steganography detection task in the real world.Currently,the application of deep learning in steganalysis is a research hotspot.However,the existing deep learning steganalysis models generally have the problem of complex model structure,which is not conducive to expansion and optimization.At the same time,the existing models cannot provide a good solution for the mismatch of embedding rate and steganographic algorithm in actual steganography detection tasks.In order to improve the performance of adaptive steganalysis networks,selection channel technology is very useful.However,reliable embedding probability estimation methods and the application of selected channel technology in new network structures still need further research.In order to solve these problems,we focus on three parts below:(1)In order to control the network scale and suppress image content at the same time,so as to better learn low-frequency steganographic noise signals,a color image adaptive steganalysis model based on multifrequency residual analysis is proposed.Most of the current deep learningsteganalyzers use wider or deeper designs to improve detection performance.However,an overly complex network will increase the training cost,which is not conducive to its expansion and optimization.In addition,steganalysis pays more attention to the high-frequency information corresponding to the image texture.However,a network with deeper convolutional layers is more likely to learn low-frequency information corresponding to the image content,which is inconsistent with the goal of steganalysis.In order to solve these problems,we propose a multi-frequency residual deep convolutional neural network MFRNet for the steganalysis of color images.The model learns different frequency components of steganographic noise in multiple ways by designing columns of different depths.In addition,by designing residual basic blocks of different residual shortcuts,it can calculate steganographic noise residuals of different scales at the same time.So the interference of image content is effectively suppressed and effectively reduces the influence of steganographic algorithm mismatch and embedding rate mismatch.Simulation results on the PPG-LIRMM-COLOR color image dataset show that the proposed MFRNet outperforms the state-of-the-art model WISERNet,especially when detecting images with a small steganography rate of 0.1 and 0.2 bpc,the detection accuracy is improved by nearly 20%.(2)To improve the performance of the selection channel technology in the real world,a reliable embedding probability estimation method is studied.Applying the selection channel technology is one of the effective methods to improve the performance of adaptive steganalysis algorithms.Many models now use the ground-truth embedding probability maps generated by prior knowledge directly to weight the feature maps of the convolutional layers.However,the computation of the ground-truth probability maps relies on a specific steganography strategy and embedded payload,which cannot be obtained in practical steganography detection tasks.To overcome this difficulty,this thesis firstly improves the existing embedding probability estimation method based on saliency detection and further proposes an embedding probability estimation algorithm based on image texture recognition.This method introduces the LBP operator into embedding probability estimation for the first time.Moreover,as a nonmachine learning method,it does not need large-scale datasets for training.Simulation results show that the proposed method can better cope with the impact of embedding rate mismatch than existing methods,especially for the images with small payloads.(3)The application of existing selection channel technology usually follows the method proposed in YeNet and uses the embedding probability maps to weight the output feature maps of the first convolutional layer to emphasize the texture information of the image.However,this method is not perfect.Firstly,the calculation of the real embedding probability map depends on the prior conditions;secondly,for existing steganalysis models with deeper and wider convolutional layers,or new network architectures with multi-channel,it is unreasonable.When training,the embedding probability information will gradually be lost as the network level deepens.Therefore,this thesis proposes an application method of selection channel technology in adaptive steganalysis algorithms.This method uses the estimated probability maps to replace the ground-truth embedding probability maps,and performs adaptive processing on them.Then performs multiple fusions with the feature maps and performs multiple weightings at multiple levels of the network,to ensure that the embedded probability information is not lost and promotes the network performance.The simulation results show that the estimated probability map is used to select the channel technique,which can further improve the model detection performance,and the method has better adaptability with deeper convolutional layers and multi-channel networks.