Research On Comprehensive Evaluation And Fusion Of Steganalysis Techniques

As an important technical tool for defending the information security, image steganographyand steganalysis have become very attractive hotspots to researchers all over the world. Imagesteganography is a technique which securely transmits the messages hidden into an innocuouscover media using the data redundancy. Steganalysis, as the opposite technology againststeganography, aims at detecting, extracting, restoring and destroying the secret messagesembedded into the cover images.So far most work of the research about image steganalysis focuses on the design ofsteganalytic algorithms. While little work refers to the comprehensive evaluation of steganalyticalgorithms. Nevertheless, reasonable comprehensive evaluation of steganalytic algorithms ishelpful in choosing the best steganalysis method aimed to different requirements of diverseapplied fields, improving steganalysis algorithms and designing steganalysis systems to conformto practical applications. By applying information fusion technology, colligating performances ofdifferent features and classifiers can improve steganalysis algorithms. Consequently, it issignificant to make a research on comprehensive evaluation of image steganalysis algorithmsand fusion steganalysis algorithms.This dissertation firstly introduces the basic theories of comprehensive evalution ofsteganalysis algorithms and fusion steganalysis, based on which, some algorithms are proposed.The contributions obtained in this thesis can be summarized as follows:1. Aimed at different requirements of different applications of steganalysis, a method ofcomprehensive evaluation of steganalysis algorithms based on combined weighting and atechnique for order preference by similarity with ideal solution (TOPSIS) is proposed. Themethod contains five performance indexes: true positive rate, false positive rate, reliability,detection error and detection speed. The entropy weight is used to obtain weights. Taking intoaccount defects of the entropy weight and the estimator’s experience and intent, the analytichierarchy process (AHP) is used to give subjective weights. The final weights are obtained withan assembled weighting method. Evaluation and comparison of steganalysis performance of thealgorithms are implemented using the TOPSIS algorithm. Experimental results show that thisalgorithm can not only be used to choose the best steganalysis method for different requirementsof each index in different applications, but also instruct to improve steganalysis algorithms.2. Steganalysis based on feature fusion: Firstly based on analyzing the validity of thefeatures extracted with the steganalysis algorithms,a weighted feature fusion steganalysis methodin which the average Bhattacharyya distance of the features is used as the weight is proposed.Experimental results indicate that the classification accuracy is greatly improved by the noveladdition of feature fusion,. Secondly a steganalysis algorithm based on multi-domain featurefusion aimed at the additive noise embedding model is proposed. This algorithm extracts thehistogram characteristic function center of mass, the moment of the histogram characteristicfunction(CF), the moment of empirical matrix diagonal projection histogram, the moment ofcharacteristic function of empirical matrix diagonal projection histogram and the middle value of projection histogram features from the spatial domain of the original image and the prediction-error image; and then the CF moments of wavelet subband coefficients’ histogram extractedfrom the original image and prediction-error image are to be as the transform domain classiferfeatures. Then the K-L transform is used for feature fusion. Experimental results show that theproposed method exhibits excellent performance3. A JPEG steganography identification method is proposed on the basis of fuzzy integralmulti-classifier fusion aimed to remedying the deficiency of the singal classifer used insteganalysis. This algorithm applies features extracted by different methods to construct SVMmulti-classifer detection systems, then applyies fuzzy integral to fuse the results of differentmulti-classifier detection systems. Experimental results show that the proposed method throughthe novel addition of the classifier fusion step improves the classification accuracy comparedwith the individual steganalysis system.Finally, the research work for this thesis is summarized and the further research topics anddirections in the future of blind detection methods, performance comprehensive evaluation ofsteganalysis algorithms and fusion steganalysis are discussed.