Image Steganography Based On Sparse Representation

Internet provides abundant information for the people. However, it also brings a lot of issues of information security. In Sparse domain steganography, we obtain the representation coefficients of image blocks with an over complete dictionary that can fully represent basic structures in images and hide secret bits in the coefficients. In this way, the semantic contents change for convey secret information. Thus sparse domain steganography has better performance when test by steganalysis. The main work of this thesis is as follows:(1) Research on dictionaries for morphological componentsWe propose sparse domain steganography based on morphological component for grayscale images. Two dictionaries are combined to represent an image sparely——one dictionary for piecewise smooth (cartoon-like) parts and the other for textures. Both of the dictionaries are highly inefficient in representing the other content. Thus the two contents will correspond to different sparse coefficients. When embedding in sparse domain, we give top priority to coefficients of textures. We present two ways to construct these two kinds of dictionaries in our work, dictionaries using mathematical models as well as dictionaries wisely learned by K-SVD algorithm.(2) Research on the relationship between local complexity and security of steganographyWe give the reason why it is difficult to use variance and entropy for complex image blocks selection in sparse domain steganography. And propose a novel concept, sparsity, to solve this problem. Image blocks used to hide secret information are selected by sparsity of their sparse decomposition coefficients. High sparsity means less nonzero entries in the coefficient vector. We sort image blocks in ascending order according to their sparsity. Low sparsity blocks have higher priority to convey secret information. We prove that blocks with more nonzero elements in their decomposition coefficients are more complex. Consequently, image blocks of high local complexity are utilized. Besides, the embedding rule in sparse domain will not change the sparsity of decomposition coefficients of blocks, which guarantees the block consistency in the extraction process. Experimental results demonstrate that our algorithm produces stego images with higher visual quality and has far better anti-detection performance than other sparse domain methods.(3) Research on sparse domain steganography in color imagesWe introduce a framework based on RGB channels in color images. We use the high correlation between the three RGB bands to transmit decomposition path. We can assume that the structural elements found in a band will also be present in the other two. Based on this idea, we calculate the decomposition path on a color band and then use it to decompose the other two. We only hide data in these two bands. The band not changed will be used to obtain decomposition path by the decoder. And the secret message in the other two bands will be recovered correctly. We apply the algorithm based on morphological component and the algorithm based on local sparsity to color image successfully.