Research On Steganalysis Of Adaptive Image Steganography

As one of the important research topics in the multimedia information security community,digital image steganography and steganalysis have drawn extensive attention from more andmore researchers. Experimental results show that different regions of cover image usually havedifferent capacities for hiding the secret message. Since the statistical characteristics of imagetexture regions are complex and are not sensitive to the secret message embedding. Thus, it isvery suitable for concealing secret messages. Adaptive image steganography utilize thedifferences of statistical characteristics between different image regions, and embeds secret bitsinto the texture regions adaptively. As a result, the security is enhanced and adaptivesteganography is difficult to be detected by traditional steganalysis algorithms. Consequently, itis of immense significance to make researches on the steganalysis of adaptive imagesteganography.In this dissertation, the image source is modeled as a local stationary Markov source. Basedon the analysis of the influences of the hidden message on the statistical characteristics of imageand the statistical differences between texture regions and flat regions, the thesis focuses on thestudy of image steganalytic algorithms of adaptive steganography. Two kinds of specificsteganalytic methods and one universal steganalytic method are proposed. The maincontributions of thesis are summarized as follows:1. The statistical characteristics of the image and the steganalytic methods for adaptiveimage steganography are discussed. The statistical differences of gray histogram, pixel-valuedifferencing histogram and gray co-occurrence matrices between cover image and stego imageare analyzed. The analytic results indicate that images with different content present differentstatistical features. Under the same condition, the statistical changes are more evident in flatimages comparing with texture images after data embedding. Finally, the state-of-art ofsteganalytic algorithms for adaptive image steganography is outlined.2. To propose a specific steganalytic method against AE-LSB (Adaptive Edges with LeastSignificant Bit, AE-LSB). AE-LSB exploits the difference value of two consecutive pixels toestimate how many secret bits will be embedded into two pixels. The range of difference valuesis divided into three levels, which decide the bit number needed to be embedded into the pixels.Before and after data embedding, the difference value of two consecutive pixels belongs to thesame level. In the covert communication, the secret message is encrypted and becomes a randomsequence of0and1. Thus, the difference values of the two-pixel blocks will be distributed in therange of some level homogeneously after data embedding. The bin grouping happens in the pixel value differencing histogram of stego image. Based on the bin grouping effect, one-dimensionclassification feature is extract from the pixel value differencing histogram of image. Extensiveexperimental results indicate that our proposed scheme has superior performance compared withthe prior steganalysis algorithms.3. To propose a specific steganalytic method against EA-LSBMR (Edge Adaptivesteganography based on Least Significant Bit Matching Revisited). EA-LSBMR selects theembedding regions according to the size of secret message and the difference value of twoconsecutive pixels in cover image. The pixel units need readjust when the differences less thanthe threshold after embedding. For the readjustment introduces abnormal fluctuations to thepixel-value differencing histogram of the stego image, one-dimension feature is extractedthrough analyzing the differences of the pixel-value differencing histograms between the coverand stego images. Extensive experimental results show that the proposed steganalytic algorithmcan effectively defeat the EA-LSBMR steganography. The method that can estimate theembedding rate as well as the length of secret messages is proposed. The order of magnitude ofprediction error can remain around10-2measured by the Mean Absolute Error, Inter-quartileRange and Standard Deviation. The order of magnitude of median absolute difference can evenremain at10-3.4. To propose a blind detection method of adaptive steganography based on local lineartransform (LLT). First, an image is filtered using different LLT mask to obtain the residue ofnoise, and then the co-occurrence matrix is extracted from the residue of noise. Further, theco-occurrence matrix and the feature extracted from difference histogram are combined as thedistinguishing feature. Finally, support vector machine is used for classification. Experimentalresults on BOWS2and NRCS image database show that the proposed method can effectivelydetect typical adaptive steganography.Finally, the research work in this dissertation is summarized and the further research topicsof adaptive steganography and steganalysis are discussed.