Research On Steganalysis Based On Image Local Complexity

With the status of information security becoming more and more important, research oninformation hiding is taken by many research institutions all over the world. By keeping up withthe new and advanced techniques in the research of media security, This thesis focuses on thedetection techniques of steganalysis. Several practical image steganalytic algorithms areproposed. The image source is modeled as a local stationary Markov source. Based on theanalysis of the different changes of the statistical features of the images with different contentscaused by secret message embedding, three detection methods based on image local complexityare proposed. The main contributions of this thesis are summarized as follows:1. Research on the statistical features of cover and stego images. On the self-built databasewith single image content, we investigate the impact of LSB matching steganography onstatistical characteristics of natural images. In addition, taking some representative steganalyticfeatures for example, we analyze the relationship between the steganalytic features and imagecontent and come to the conclusion that the statistical changes are more evident in flat regionsafter embedding, which provides a theoretic basis for researching image content basedsteganalysis techniques and proposing new and reliable steganalytic algorithms.2. A steganalysis of LSB matching based on deviation degree histogram. We model theLSB matching embedding as additive noiseand analyze the correlation of local pixels. Deviationdegree is defined to reflect the local complexity.56features are extracted from the deviationdegree histogram of flat region to build a two-class classifier. Experimental results demonstratethat the proposed method is efficient to the LSB matching steganography on uncompressedimages and significantly outperforms prior arts.3. A steganalysis method based on the run length histogram of flat region. Firstly, thequad-tree segmentation based on the deviation is used to get the areas which are sensitive tomessage embedding, then, features based on run length histogram are extracted and optimized,finally, support vector machine is employed to discriminate covers and stegos. Experimentalresults show that the new steganalyzer outperforms the existing methods in detection accuracy.4. A Blockwise joint judegement steganalysis method using Mean Shift image segmentation.We decompose images into small sub-images, categorize these sub-images based on image localcomplexity, train a classifier for each category and assign different weights to each category. Thefinal detection result of a whole image is obtained by weighted fusion of the results of its sub-images. Experimental results indicate that this approach outperforms the representativesteganalysis algorithms.Finally, we summarize our research work and discuss further research topics and futureresearch directions.