Research On Adaptive Steganalysis Model Based On Deep Learning

Steganography is conducive to the safe transmission of information,but the abuse of steganography will bring many potential security problems.And steganalysis plays a vital role in preventing the abuse of steganography.This thesis researches the adaptive steganalysis model combined with deep learning technology from three aspects.In recent years,many models based on Convolutional Neural Network(CNN)have been proposed.Firstly,in the current models,to enhance steganographic noise,the image is generally pre-processed to generate the image residual.However,the generated image residual is directly used for the feature extraction of the subsequent network,and the residual processing is simple.Secondly,the adaptive steganography algorithm realizes the embedding of information in the texture-rich area of the image.However,in most models,the residual elements are input to the subsequent network with the same importance,thus,it is difficult to learn the key features.Finally,in practical applications,the amount of information embedded in the detected images is generally unknown and then the payload mismatch problem may occur when the trained steganalysis model is used for detection.This thesis studies the above three issues,the contents are as follows:1.To enrich the extraction of residual direction feature,this thesis proposes an image steganalysis model called ResD-CNN combined with the directionality of residuals.To introduce the residual directionality,residuals are rearranged in the direction of minor-diagonal prior to the feature extraction,so as to obtain the residual minor-diagonal directionality.In addition,Local Binary Pattern is applied to the residual map,so as to obtain the correlation between each element in the residual map and its multi-directional adjacent elements.Current steganalysis models contain many learning parameters,and a large number of convolution operations are applied to feature extraction,which requires more computing resources and affects the efficiency of the model.Therefore,the proposed model ResD-CNN reduces convolution computation and the number of learning parameters,so as to improve the efficiency of the model.ResD-CNN has 65,000 learning parameters.For WOW adaptive steganography algorithm with a payload of 0.4bpp,the detection accuracy of ResD-CNN is 0.77,which is better than that of Qian-Net model.2.To improve the key feature extraction ability of the model,this thesis proposes an image steganalysis model called IAS-DSCCNN,which combined with diversified selection channels.The adaptive steganography algorithm calculates the embedding distortion of each pixel and then selects the appropriate area,usually the texture-rich area,to embed information.According to this characteristic,the embedding probability map generated according to the steganography algorithm and payload is combined with CNN to introduce the selection channel,so as to enhance the residual of the area with a high probability of embedding information.At the same time,according to the distribution of the elements in the embedding probability map,the discrimination of each element in the element concentration area is increased,so that the residual elements are input into the subsequent network with different importance,which promotes the network learn the key features.To enrich the application of the selection channel,this thesis extracts the maximum values of adjacent elements in different directions to generate horizontal and vertical maximum residual enhancement maps,and applies them to residual respectively,so as to improve the statistical modeling ability of the network.For S-UNIWARD adaptive steganography algorithm with a payload of 0.4bpp,the detection accuracy of IAS-DSCCNN is 0.738.IAS-DSCCNN has similar performance with Yedroudj-Net,and compared with most steganalysis models such as Yedroudj-Net,IAS-DSCCNN has less residual extraction and convolution operations,thus,it has advantages in the condition of limited computing and storage resources.3.In the current steganalysis models,CNN is generally applied to extract the global features of the image and then learn the differences to distinguish between natural image and steganographic image.However,the global features are closely related to image properties,such as payload.As the steganalysis model is trained with the dataset containing a specific payload of steganographic images,it performs well when detecting images of the same payload,but mostly works badly if not.In practical applications,the amount of information hidden in the detected image is generally unknown.Therefore,the problem of payload mismatch occurs.To reduce the requirement of the model for the consistency of image property,this thesis proposes a detection method based on local block difference analysis,which realizes the detection by analyzing the feature change pattern between different local blocks of the image,so as to reduce the influence of the payload mismatch on the detection performance of the model.At the same time,this thesis introduces long short-term memory network technique into image steganalysis and proposes a new steganalysis network architecture called ILDA-LSTM based on the image local difference analysis,which realizes the steganalysis by analyzing the changes among the elements in the spatial sequence composed of local image features.The simulation results indicate that,for WOW adaptive steganography algorithm,the detection performance decay rate of ILDA-LSTM is slower than that of the steganalysis model based on CNN,thus,the local difference analysis model architecture can effectively reduce the impact of payload mismatch on model detection performance.