| Steganography is an information hiding technology that conceals secret information in a carrier such as an image,so that the information will not be discovered during the transmission process,and the security of the information is protected.However,in addition to being used in copyright protection and covert communication scenarios,steganography has also been abused by some criminals in illegal activities that are harmful to social security.Therefore,the legal use of steganography needs to be effectively regulated.Steganalysis is a technique for detecting whether steganographic information is contained in a carrier.Steganographic features are usually extracted based on the fact that steganography will change the correlation between neighboring pixels of an image,and then use a classifier to classify these features to determine whether the information has been hidden.Classic steganalysis algorithms such as Spatial Rich Model(SRM)and its derivative models have a good detection effect on common steganography algorithms.However,with the continuous development of steganalysis technology,steganography algorithms are also constantly improved.The adaptive steganography algorithm has made fewer changes to the image,and it is difficult for common steganalysis algorithms to detect them.In addition,SRM and other steganalysis models have the problems of high feature dimension and low algorithm efficiency.This dissertation studies the difficulties in steganalysis.The main contributions are as follows:1.A new feature is introduced to describe the steganalysis noise and reduce the feature dimension of steganalysis algorithm.Local binary pattern(LBP)is a classical image texture description operator,and it can reflect the relationship between neighboring pixels.Therefore,this dissertation introduces LBP into steganalysis method,and proposes a steganalysis model called demsLTP.The model uses derivative filters to obtain the residuals,and extracts LBP features with multi-size mixed resolution,and uses the modified Fisher score to screen the better performing LBP features.Simulation results show that demsLTP has a high detection accuracy rate similar to SRM while maintaining high algorithm efficiency.2.This dissertation combines image quality measures(IQM)and SRM technology and reduce the feature dimension of steganalysis algorithm.The SRM method uses linear and non-linear filters to obtain various types of residuals to represent the relationship between adjacent pixels.Although it has a good detection accuracy,its feature dimension is too high,and the detection effect is affected by the differences in the natural attributes of the source.Based on the idea of IQM,this dissertation proposes a steganalysis model called SIQMM,which takes the distance between the image and its filtered version as the detection feature.SIQMM selects the TV filter and some filters from SRM to filter the image to eliminate possible steganographic noise in the image,and then designs a quality measurement algorithm to calculate the distance between the original image and the filtered image.This distance is used as a feature of steganalysis.The number of features used by SQIMM is only 1/50 of that used by SRM,but it can still achieve the same accuracy as SRM.3.This dissertation introduces the idea of selecting channels and achieve adaptive steganalysis based on deep learning.Steganographic analysis methods based on deep learning have gradually become a research hotspot.The features used in the steganalysis methods represented by SRM are all from artificial design,although they have achieved good results,the development of the algorithm is affected by many factors such as expert experience and design cost.The feature self-extraction capability of deep networks solves this problem.However,existing networks have the problem of large parameter scales and difficulty in detecting adaptive steganography algorithms.This dissertation proposes a steganographic analysis model ASCNN based on convolutional neural networks.ASCNN has only 5 convolutional layers and has high efficiency.At the same time,in view of the feature that the adaptive steganography algorithm always embeds information into regions with smaller changes to the image,ASCNN emphasizes regional differences when detecting.In other words,the embedding probability graph of the image is calculated to weigh the residual of the region with high steganography probability.This enables the steganalysis model to focus on these key features,and enhances the detection ability of the model for the adaptive steganography algorithms.Simulation results show that ASCNN has good detection ability for many classic adaptive steganography algorithms. |
