Research On Key Issues Of Steganalysis Of Content-adaptive JPEG Steganography

Content-adaptive steganography based on the framework of “distortion function + coding” is the most successful form of modern steganographic schemes for digital images,the embedding changes are performed in regions that are difficult to describe using statistical models,such as complex textures and noisy areas,which brings great difficulties and challenges to steganalysis.This thesis focuses on the steganalysis of content-adaptive JPEG steganography,mainly discusses the researches on the extraction of low dimensional sensitive features,the reliable detection of specific steganographic scheme,the exploit of change probabilities,and the optimization of ensemble classifier.The main works and research results are as follows:1.The theoretical and practical significance of the research on digital image steganography and steganalysis are introduced,the research progress and development at home and abroad are reviewed,and several key issues concerning the research on the steganalysis of content-adaptive JPEG steganography are stated.2.The extraction of low dimensional sensitive features.For the blindness in the construction of filters which results in the features having high dimension but low sensitivity,a new feature extraction method based on Gauss partial derivative filter bank is proposed.Considering the embedding characteristic of content-adaptive steganographic schemes,the proposed method constructs the filter bank using Gauss partial derivatives with multiple scales and orders,and histograms of filtered subimages are extracted as steganalysis features.Gauss partial derivative filter bank can represent texture and edge information in multiple orientations with less computation load than conventional methods,which is beneficial in the extraction of low dimensional sensitive features.Experimental results indicate that the proposed features achieve competitive performance with lower dimension compared with the prior art.3.The reliable detection of specific steganographic scheme.For the flaw that the existing methods fail to consider the characteristics of perturbed quantization steganographic schemes,and there is redundancy in the features,a new feature extraction method based on possible change modes is proposed.Considering the initial screening of changeable coefficients using selection rule in these schemes,the proposed method identifies possible change modes using the relationship between changeable coefficients and the second quantization steps,and extracts features based on the possible change statistical samples of the existing feature extraction source,the obtained features have less redundancy and are thus more sensitive to these schemes.Experimental results demonstrate the validity of the proposed method and show that the combination of multi-order statistical features can further improve the detection performance.4.The exploit of change probabilities.A feature extraction method based on the weight allocation of filtered coefficients is proposed.Considering the coefficients in the embedding domain have different change probabilities,the proposed method expresses the filtered coefficient using correlation DCT coefficients by analyzing their relationships,and calculates the maximum change probability of correlation DCT coefficients as the weight of corresponding filtered coefficient,the final features are generated by accumulating the weights of filtered coefficients with equal value.The proposed method enhances the description ability of features for embedding changes,and therefore have better detection performance.Experimental results show that the proposed method can improve the detection performance of existing features.In addition,it has advantages whether the payload is known or unknown compared with SCA method.5.The optimization of ensemble classifier.For the problem that the existing ensemble classifier ignores the relativity of feature components and the performance difference of base classifiers,a steganalysis method based on the genetic and Pareto algorithm is proposed.The genetic algorithm with good global search ability is employed to construct the feature subspace,in which feature subset is recognized as the basis selection unit,the separability and diversity of feature subspace are thus ensured.The Pareto algorithm with dual objective optimization is employed to optimize the obtained base classifiers,the fusion result and the number of classifiers are all considered.Experimental results show that the proposed method can improve the detection performance of ensemble classifier.