| With the popularization of the Internet and the rapid development of information technology,steganography and steganalysis have attracted widespread attention in the field of information security.The main purpose of steganography is to hide secret information in other information carriers and transmit it;correspondingly,steganalysis is to successfully detect hidden information.Traditional steganalysis methods generally include two steps: feature extraction and classifier training.In these two steps,feature extraction is the key,which plays a decisive role in detection performance.In the past decade,feature extraction mainly relied on hand-crafted features.The basic idea is to find feature statistics that have obvious differences in the information carriers before and after the steganography operation.In other words,traditional steganalysis techniques need to rely on a large amount of prior knowledge to design an effective way to extract features,and this process requires a lot of human resources and material resources.With the revival of artificial intelligence technology represented by deep learning in recent years,steganalysis methods based on deep learning have been proposed.Compared with traditional steganalysis techniques,steganalysis methods based on deep learning can directly obtain stegographic features through a large amount of training data.However,while the steganalysis methods based on deep learning brings new development opportunities,it also faces many challenges.For example,the steganalysis features obtained by the deep learning model are not discriminative,which makes it difficult to fully detect the steganography information;the existing steganalysis models based on deep learning are not flexible enough to be suitable for different information carriers.To this end,this thesis focuses on the spatial image steganalysis method based on deep learning.Starting from two aspects of steganographic features and steganographic models.It solves the stubborn problems that the features are not discriminative and the model is not flexible,which bother the deep learning-based spatial image steganalysis technology.The main contributions of this thesis are as follows:First,in the task of image steganalysis,it is necessary to extract the residuals of images,where the positive and negative of the residual value correspond to the positive and negative of the steganographic information.For the model,both the positive and the negative play the same role,which is different from the traditional computer vision.Therefore,for steganalysis tasks,an activation function with a symmetric structure should be selected to ensure that the model has the same judgment on positive and negative information.Based on this,this thesis analyzes and compares the influence of activation functions such as Gaussian kernel function and linear rectification function on the steganalysis model,and finds the activation function that is most suitable for the steganalysis model,which is linear truncation function,through experimental comparison.In order to verify the performance of steganalysis based on linear truncation function,three common steganalysis algorithms are embedded into BOSSbase,BOWS2 and AUG datasets to construct datasets we need,and the steganalysis model based on linear rectification function is used as the benchmark model for experimental comparison.Compared with other activation function methods in steganalysis,our method has a lower detection error rate and a faster convergence rate.Second,in actual application requirements,the carrier of image steganography is of arbitrary resolution,while the existing deep steganography model can only achieve singleresolution training and single-resolution detection.In view of the practical problem that the deep steganography model cannot be applied to image carriers of different resolutions,a steganalysis method of multi-resolution spatial image based on selective channel perception is proposed in this thesis,where the spatial pyramid pooling is introduced into the steganalysis model.By introducing spatial pyramid pooling,images of different resolutions are pooled into the same size.In addition,the selected channel is used as the prior knowledge of the secret images and integrated into the deep neural network.So that the deep neural network can analyze the secret images more accurately to preliminarily screen out image areas with high probability of steganography.The experiment carried out a comparative analysis on the classic spatial enrichment model,maximum spatial enrichment model,and spatial image steganalysis based on linear truncation function.The results show the proposed multi-resolution spatial image steganalysis method based on selective channel perception can greatly improve the detection accuracy of the steganalysis model under different embedding rates. |
