Hypergraph And Sparse Representation Research In Steganography

With the rapid development of Internet, the issue of information security has increasingly become important and outstanding. As the main measure in secret communication and protection of intellectual property, information hiding technology has been widely researched and used. And as one of the important branch in information hiding technology, steganography can be used in personal privacy, business, military departments and the secret communication between governments.In recent years, hypergraph and sparse representation are widely used in information retrieval, data mining, image denoising and classification. This paper explores hypergraph and sparse representation, then applies them into stegangraphy. All three aspects are as follows:(1) This part proposes a hypergraph-based steganographic method. In this method, pixels are divided into groups. Hyperedges are constructed by exchangeable relationship between pixels. A hypergraph that is the set of hyperedges is used to represent the cover image. By looking for hypergraph matching using a local optimal greedy algorithm, the secret message is embedded. Experiments show that comparing with the stegangraphic algorithm based on graph theory, the proposed method requires less modification of pixels and performs better in visual quality. And it is can better resist steganalysis than other spatial domain methods.(2) Decompose cover image into blocks with redundant dictionary and embed secret messages into the decomposition coefficients in the sparse domain. In this chapter, we construct binary redundant dictionaries with different learning algorithms and compare the different redundant dictionaries’ performance in the process of decomposing images, embedding secret messages, reconstructing images. Experimental results show that learning dictionaries performs better in embedding capacity and anti-steganalysis ability than manual dictionary and dictionary based on mathematic model.(3) According to the defect of quantizing error in the binaryzation of the redundant dictionary, non-negative constraint is introduced in the construction of dictionaries.In this chapter, dictionary atoms are created with non-negative learning algorithms. And the nonnegative learning dictionaries and general learning the dictionaries are compared in three aspects, including embedding capacity, anti- steganalysis ability and anti-steganalysis ability when embedding rule is improved. Experimental show that, comparing with the general learning dictionary, nonnegative learning dictionaries’embedding capacity dropped slightly, but the ability in resisting steganalysis is improved, especially used the improved embedding rule. Overall, the best negative learning dictionary-ICA dictionary is more suitable to steganography in the sparse domain.