| Steganography and steganalysis are necessary to research focusing on information hiding.However,the abuse of the steganography algorithm causes many security risks.For example,criminal organizations pass illegal information through steganography to complete covert communication.Image steganography is a technology that achieves the purpose of covert communication by embedding secret information.Due to the particularity of steganography algorithm,it may be used in illegal fields.Therefore,the research of the steganalysis algorithm has essential research significance and practical value for maintaining information security.The existing steganalysis algorithms based on deep learning have high redundancy and loss problems.In this paper,the applicability of steganalysis is enhanced through FY-BOSSBase composite data set.At the same time,the Vision Transformer model is combined with the steganalysis model.A new convolutional steganalysis network CTS is proposed by adding an attention mechanism.CTS network improves classification accuracy while reducing model parameters and has good detection performance.The main contents of this paper are as follows:1.Generate compound steganographic data based on Fisher-Yates random scrambling algorithm.A verification framework for steganalysis in the image domain based on random distribution is designed.First,expand the BOSSBase v1.01 dataset based on the Fisher-Yates algorithm to generate the FY-BOSSBase composite dataset.Secondly,three steganography algorithms,ns F5,J＿UNIWARD,and UERD,organize the FY-BOSSBase combined dataset.Then,the principle of SRNet steganalysis is analyzed in depth,and the loss iteration of the SRNet network is carried out using cross-entropy.Finally,realize random distribution steganography analysis in the image domain to verify the model’s validity in the image domain.2.Through the combination of ConVit and CNN.A new network framework CMVT(Convolution meet Vision transformer)is proposed.CMVT utilizes the translation invariance and local variability of a convolution neural network.And GPSA in ConVit focuses on local information to obtain a larger image perception area.This improves the steganographic model detection accuracy and reduces the model’s cross-entropy loss.3.A convolution steganalysis network CTS(Convolution Transformer Steganalysis)is proposed for image steganalysis.The CTS network first adds a CBAM attention mechanism in the preprocessing stage to capture the local features of the steganography in the image.Secondly,in the feature extraction stage,we focus on the global correlation of the image through convolution projection to effectively extract the noise residuals.Finally,the loss function of different distance functions is compared to adapt to the image changes before and after the secret information is embedded.CTS network performs satisfactorily on fixed-size data sets and improves detection accuracy. |
