Research On Image Adaptive Steganalysis Technology Based On Deep Learning And Local Difference

As an important means of covert communication,image steganography has been widely used.At the same time,security incidents that criminals maliciously use steganography to hide malicious code or secret information in image carriers also occur frequently.Therefore,using steganalysis technology to detect whether an image contains secret information is of great significance for protecting digital content security.In recent years,researchers have proposed a variety of steganalysis models based on convolutional neural networks,which effectively improve the detection ability of steganalysis models.However,there are still problems such as insufficient detection performance and significant performance degradation in the case of payload mismatch.It can be seen that it is of great practical significance to design a steganalysis model with a new architecture,improve the detection ability of the model,and promote the practicality of steganalysis technology.In view of the above problems,this dissertation carries out research from the following aspects:(1)Research on steganalysis model based on local difference analysisAiming at the problem that the current steganalysis model based on global features is easily affected by the attributes of the image itself,which leads to the degradation of detection performance in the case of payload mismatch,this thesis proposes a new steganalysis model LDANet based on local difference analysis.The model determines whether an image is a steganographic image by analyzing the difference of spatial smoothness and feature attributes of different local blocks of the image,which improves the detection performance of spatial domain steganalysis model.Firstly,LDANet uses the convolutional neural network based on residual structure to extract the steganographic features of local blocks and compose the feature sequence.Then,the Long Short-Term Memory(LSTM)network is used to learn the spatial smoothness difference of the feature sequence according to the Hilbert curve order.Finally,the smoothness difference features and the statistical features of the sequence are fused to classify the steganographic images.Different from the loss function design method that only relies on cross-entropy,LDANet designs a loss function including contrastive loss,local block classification loss and overall network classification loss,which guides the network to better extract the local features of the image.Experimental results show that LDANet has the highest accuracy of 92.05%when detecting the steganographic images generated by the WOW steganographic algorithm with a payload of 0.4 bpp.The detection performance of LDANet is also better than that of other state-of-the-art steganalysis networks.Compared with YeNet,SiaStegNet and SRNet,the detection accuracy of LDANet is increased by 11.78 percentage points,3.07 percentage points and 1.05 percentage points respectively.The detection performance of LDANet is also better than SRNet in the case of payload mismatch.(2)Research on steganalysis model based on denoising autoencoder and Siamese networkAiming at the problem that the current ASSAF steganalysis model based on denoising autoencoder and Siamese network has poor performance and is only suitable for JPEG domain,this dissertation proposes a steganalysis model DAES-Net based on denoising autoencoder and Siamese network.Different from the existing steganalysis models based on convolutional neural networks that directly extract the steganographic features of images for classification,DAES-Net uses the Siamese network to extract the difference features between the original image and the steganographic noise image for steganalysis detection,which extends the network architecture design idea of the steganalysis model.The preprocessing module of DAES-Net used the denoising autoencoder to remove the steganographic noise of the input image.The feature extraction module used the convolutional neural network based on residual structure and Siamese structure to extract the steganographic difference features between the input image and the denoised image.According to the characteristic that the feature difference distance between the original image and the steganographic image and the respective denoised image are different,the classification module calculates the distance measure between the features,and uses the fully connected layer to classify,so as to determine whether the image is a steganographic image.The experimental results show that the denoising encoder of DAES-Net can effectively remove the steganographic noise of the image,and the generated image not only has more detailed texture,but also outperforms the image generated by ASSAF in both visual effect and statistical analysis.DAES-Net has the highest accuracy of 91.01%when detecting the steganographic images generated by the WOW steganographic algorithm with a payload of 0.4 bpp.When detecting WOW,S-UNIWARD and HUGO steganographic algorithms,the detection performance of DAESNet is increased by 7.53 percentage points,8.85 percentage points and 7.75 percentage points respectively compared with ASSAF.