Steganalysis With High Dimensional Features And Secure Steganography In Digital Images

While digital steganography aims to embed secret message into digital media,such as digital images, digital radio and video, the purpose of steganalysis is to detectthe presence of secretly hidden data in a digital media. The reaserch in steganalysisand steganography has important academic values and realistic meaning. Thisdissertation mainly focuses on steganalysis and steganography in digital images andsummarizes the results obtained in feature extraction, classifier designing and featuredimensionality reduction. The contributions of this dissertation are listed as follows:1. Steganalysis for JPEG images based on high-dimensional features and Bayesensemble classifierThis dissertation proposes a JPEG steganalysis scheme based onhigh-dimensional features and Bayes ensemble classifier. The weak correlationbetween DCT difference coefficients may be destroyed when the datas are embeddedby steganographic algorithm. We design7850dimensional features from differenceco-occurrence matrices, and then combine the7850dimensional features fromcoefficient co-occurrence matrices to describe the statistic characteristic in JPEGimages. Then, original images and stego images are classified by ensemble classifier.Multiple subclassifiers can be generated and by introducing the Bayes theory, themultiple decisions from these subclassifiers are optimized. Finally, the optimaldecision is given. Comparing with the stat-of-the-art steganalysis scheme,2%improvement can be obtained by merging high-dimensional features and Bayestheory.2. Steganalysis for color JPEG images based on ensemble proportion trainingWe propose a new steganalytic scheme of color JPEG images based on YCbCrcolor space. The features of proposed scheme include intra-channel features and inter-channel features. The intra-channel features, including Markov features,extended DCT features and co-occurrence matrices features, are extracted in Ychannel and can capture effectively the dependency among DCT coefficients in Ychannel, while the inter-channel features are extracted in difference planes betweensample planes from Y channel and CbCr channels, which can capture the dependencybetween channels. In classifying processing, we respectively use intra-channel andinter-channel features to train subclassifiers. The optimal ensemble decisions can bederived by adjusting the proportion of two kinds of sub-classifier. The proposedscheme has good detection for color JPEG image and the performance outperformssome state-of-the-art feature sets.3. Spatial steganalysis based on local textural pattern and double dimensionalityreductionA spatial steganalysis scheme is proposed based on local textural feature anddouble dimensionality reduction. The textural distribution of original images may bechanged by adaptive steganography. The local textural pattern can be obtained bycomparing the pixel values with the neighbors’ value in residual images. Thehigh-dimensional textural features are formed by combining all local textural patterns.Then, principle component analysis is used to perform double dimensionalityreduction for high-dimensional textural features. The correlation in high-dimensionaltextural features obtained in the same filter can be eliminated by first dimensionalityreduction, while the correlation from different filters can be also eliminated by seconddimensionality reduction. Finally, a textural feature set with low dimensional isproposed and can effectively performed steganalysis by combining the classifier.4. JPEG secure steganographic scheme based on new distortion functionIn order to improve the secure of JPEG steganography, by considering thestatistic features in steganalysis schemes, we design a new distortion function bycombining intra-block residual, inter-block residual and coefficient values in DCTcoefficients. The optimal parameters in new distortion function are found byexhaustive search. The secret messages are embedded into JPEG images by usingsyndrome trellis coding (STC) and distortions of DCT coefficients, meanwhile, the total distortion of stego image can keep minimal. Comparing with some existingschemes, the proposed scheme has higher security performance and low computingcomplexity.This dissertation has an extensive study in steganalysis in digital images, mainlyfocusing on high-dimensional feature, classifier and feature dimensionality reduction,and also considers the improvement in secure steganography. These works areimportant in helping the development between steganography and steganalysis.