Research On Steganalysis In Digital Images

Steganography and steganalysis are two conflicting subjects.The steganography is used to covert communication with multimedia data and,on the contrary,the steganalysis is the art to detect the presence of steganography.This dissertation mainly focuses on steganalysis.Combining the content-adaptive steganography schemes,this dissertation proposes several effective steganalytic schemes.Meanwhile,based on the theory of ensemble learning,we propose diverse ensemble classifier to improving the detection accuracy of classifier.The major works are listed as following:1.Low-dimensional spatial steganalysis using redistributed residualsWe propose 2929-dimensional features represented with first-order statistics.The features are combined with two parts.The first kind of feature is the original residuals formed with larger threshold to capture the steganographic changes distributed in the edge and texture region of cover.The second kind of feature is the redistributed first-order statistics.With the aid of the anti-correlation of adjacent residuals and the shifting parameter,the the residuals can be redistributed.The redistributed mechanism can enhance the diversity of feature and improve the detection accuracy.For testing the effectiveness of the proposed feature,we do the test on the conten-adaptive steganographic algorithm.Compared with previous scheme on low embedding rate,the redistributed feature is effective to existed steganographic algorithm by about 5.65%.2.Spatial steganalysis using the texture of images and the contrast of residualsA novel 2363-dimensional scheme for spatial steganalysis is proposed based on the extracted texture blocks and the contrast of residuals.In the first part,according to the mechanism of content-adaptive steganography method,the secret information is embedded in the edge and texture region of cover and the smooth region is not used for embedding the secret message.In order to extract the effective feature,the fluctuation function is defined to evaluate the complexity of cover and we can obtain the most complex area(subimages).The selected subimages are the optimal region for embedding.For capturing the steganographic changes,with the linear and non-linear filters,the diverse residual images are computed from the selected subimages and original image.Then,the contrast of residuals is projected onto an angle and the ratio is transformed into the angle.Meanwhile,the l2-norm of residuals is denoted as the corresponding weight of angle and the feature is the combination of angle and weight.The proposed feature can present the joint statistic of different residuals and the dimension of the proposed feature is linear with the threshold.Hence,this claracteristic can be used to obtain low dimensional feature.Compared with previous spatial steganalytic algorithm,for content-adaptive steganographic method,the proposed steganalytic method can improve the detection accuracy with low-dimensional feature.3.Steganalysis of JPEG images using multi-resolution decomposition and affine transformation.We propose a novel scheme for steganalysis of JPEG images with multi-resolution decomposition and affine transformation.We think the JPEG image is made up by several different resolution subimages through non-linear combination mode.Hence,the original JPEG image is decomposed and the steganographic changes can be distributed into the JPEG subimages through non-linear mode.Moreover,the steganographic changes can be exposed by discarding the smoothness component and the detection accuracy is impoved.For obtaining the multi-resolution images,the JPEG image is decompressed into spatial domain and the spatial version is decomposed into several subimages with multi-resolution tool.Then,all the subimages are compressed into JPEG domain and we can obtain several JPEG images.For any JPEG images,all the DCT(Discrete Cosine Transform)planes are considered as the individual planes and the residuals of DCT plane is computed along different directions.The ratio of different residual DCT coefficients is turned into an angle and the final feature is denoted as the combination of angle and the l2-norm of residuals.For the purpose of enhancing the diversity of feature,we use the affine transformation to rotate the projected angle and get new representation of original feature.Compared with previous scheme,on different embeddng rates,the proposed scheme can improve the detection accuracy with the aid of multi-resolution decomposition and affine transformation.4.Diverse ensemble classifierWe propose a diverse ensemble classifier scheme.The ensemble classifier can be used to work with high-dimensional feature with low significantly lower training complexity.However,there are two shortages in the previous ensemble classifier.The first one is the chosen mechanism of final classifier.In the original algorithm,most base classifier is abandoned and the chosen one is the classifier owing minimum testing error.In order to prevent the proplem of overfitting and enhance the generalization ability,we use the Bagging mechanism to generate the classifier which can use fully the abandoned classifier.The Bagging can enhance the generalization ability and prevent the proplem of overfitting.The second one is the weak classification ability of single classifier.This characteristic may affect the detection ability of final classifier.In order to improve the classification ability of the final classifier,we adopt the Adaboost mechanism to enhance the detection ability of subclassifier and make the final classifier to be a more powerful classifier.The detection results of many kinds of steganalytic features show that the diverse classifier can improve the detection ability of feature.This dissertation mainly focuses on the steganalysis and do further researches on the redistributed residuals,complexity of cover,contrast of residuals,multi-resolution decomposition and diverse classifiers.