Steganalysis Method Based On Posterior Information Of Carrier

With the rapid development of network multimedia and steganography,computer users can easily achieve secret transmission of information with the natural masking of digital carrier.In practice,this technology is able to be used by illegal criminals to threaten social and national security.Steganalysis,an effective countermeasure against steganography,can detect whether a digital carrier is a cover or a stego which embodies its important value of research and application.In many cases,steganalysis accuracy is not satisfactory even under high payload conditions due to the strong disturbance of carrier signal.The classification model with a low detection accuracy cannot provide valuable conclusions in the real world.Therefore,the detection of signals with very low signal-to-noise ratio is a highly challenging task.Existing studies have shown that the properties of carrier are very important for detection and these studies have made great progress in steganalysis.However,the correlation between carrier and steganalysis accuracy has not been studied carefully so far.In this thesis,it is found that there exists significant individual differences between the masking effects of carriers.In the real environment,the masking strength not only increases the challenge of detection,but also brings the opportunity of successful detection.If inherent attribute information of the carrier is used to enhance steganalysis features or evaluate confidence of detection results,the application value of the existing steganalysis model will be enhanced.Aiming at these problems,this thesis obtains the posterior information of steganalysis from digital carriers and then makes targeted enhancement and revaluation of steganalysis features and accuracy respectively.The main work in this thesis includes the following four aspects:1.An extraction method of fusion feature for motion vector steganalysis is proposed.Based on the principle of motion vector correlation,a 16-dimensional HVDL feature is obtained by calculating the entropy means of each lowest 4 bits from the horizontal component H,the vertical component V,the direction D and the length L of motion vectors respectively.In this feature,a local range of entropy values is used as a strength description of the motion vector correlation.Based on the principle of motion vector optimality,the SAD and RSAD increments are used as the carrier information to perform calibration weights of eight motion vector pairs in the neighborhood,and the 16-dimensional motion vector feature SIC is extracted.A motion vector pair instead of an individual motion vector is used as the basic unit of feature statistics.Experiments show that the carrier information based calibration feature has a higher accuracy than traditional features,and the 32-dimensional fusion feature HVDL+SIC has a better performance than the single feature because the shortcomings of single feature are overcome.2.A method for evaluating the posterior accuracy of video motion vector steganalysis is proposed.In this thesis,the influence of motion vector steganography on coding cost is analyzed theoretically.It is demonstrated that the macroblock complexity has direct correlation with the local optimality of motion vectors.The probability distribution of macroblock complexity indicates that the macroblock complexity has a large distribution range,but most of the macroblocks have a lower complexity.Both theory and experiments show that video frames with more complex macroblocks enjoy higher detection accuracy for optimal motion vector based steganalysis features.Based on this rule,in this thesis,the posterior information value P of one video frame is calculated by the weighted summation of probability distribution of the macroblock complexity.Experiments show that revaluation of detection accuracy of motion vector steganography can be realized by the value P.Compared with the prior accuracy,the detection reliability of individual samples or test subsets can be effectively distinguished.3.A method for evaluating the posterior accuracy of spatial images steganalysis is proposed.In this thesis,adaptive convolution is proposed to obtain the low frequency signal of the spatial image,and then the difference between the original image and the obtained low frequency image is calculated as the high frequency noise of the spatial image.After the elements in the residual matrix are truncated and quantized,the second-order co-occurrence probability matrix is calculated to describe the probability distribution of the truncated quantized residuals,which is taken as 196-dimensional feature ACF.The clustering of ACF features in this thesis shows that the more concentrated distribution of co-occurrence probability,the higher detection accuracy will be got.Based on this law,the co-occurrence probability in the ACF feature is weighted and summed as a value S which is used to measure the concentration degree of the probability distribution.The posterior accuracy based on the S value indicates that it can evaluate the performance of a single sample or subset more effectively than prior accuracy.The detection accuracy of the samples,which have high S values,can even reach 100%.4.A method for evaluating the posterior accuracy of JPEG images steganalysis is proposed.In order to study the spatial correlation of JPEG images,this thesis performs local linear regression on spatial pixels of JPEG images to obtain a regression coefficient matrix.The values of coefficient matrix show that the spatial correlation is weakened significantly with the increase of pixel distance.In order to further study the correlation between pixels,this thesis selects four types of adjacent pixel pairs in JPEG spatial domain and calculates the corresponding Pearson correlation coefficients.The mean value C of the four correlation coefficients is calculated as a posterior information of the JPEG image.The value C can approximately represent the linear correlation between the spatial domain pixels.Experimental results show that the steganalysis accuracy of JPEG images has a significant linear positive correlation with value C.Compared with prior accuracy,it is more reliable to use the value C to posterior evaluation accuracy in the case of the same embedding rate(bpnzAC).The relationship between the non-zero AC coefficients and the value C indicates that posterior accuracy will enjoy a better performance of the JPEG images which have the same embedding rate of bpp.