Research On Image Steganography Without Embedding

Image steganalysis,as a technique to detect whether the image carrier has hidden secret information,has been developed in the long-term confrontation with the image steganography technology,and has become a threat to the security of image steganography.Traditional image steganography embeds secret information into carrier image by modifying the image data directly,thus,it is hard to resist the detection of image steganalysis.Due to the shortcomings of embedding-based image steganography,researchers put forward the concept of non-embedding-based image steganography,which hides secret information into the carrier image while keep the content of image unchanged.Therefore,it can effectively resist the detection of steganography analysis and improve the security of image steganography.However,state-of-the-art non-embedding-based image steganography algorithms generally have the problems of low steganography capacity and single image type(special texture image,etc.).To solve these problems,different non-embedding-based image steganography algorithms are proposed in this dissertation.Firstly,this dissertation proposes a non-embedding-based image steganography algorithm with the help of image robust hashing.According to the characteristics of image SIFT(scale-invariant feature transform)algorithm,we design and implement a robust image hash which can resist common image attacks.It increases the steganography capacity of non-embedding-based image steganography algorithm implemented by carrier selection.And the robustness of the stego image is enhanced.Furthermore,the size of local image database is reduced by establishing the quadtree index structure,and the time cost of selecting carrier image is also reduced.Secondly,this dissertation designs and implements a non-embedding-based image steganography algorithm with the help of DCGAN(deep convolutional generative adversarial network).We train DCGAN to generate stego images without embedding or modification according to secret information.And a CNN(convolutional neural network)-based extraction model is designed according to the characteristics of generation model in DCGAN to extract secret information from generated stego images.Compared with the existing non-embedding-based image steganography,the propsed algorithm has the vittures of no need to establish the carrier image database,and the generated stego images approximate natural images in the real world.In the end,the effectiveness of the proposed algorithms are proved by experiments,and the performance of the algorithms are analyzed from different perspectives.Experimental results show that the proposed algorithms can use natural images or approximate natural images as stegos.Compared with existing non-embedding-based image steganography algorithms,the steganography capacity of algorithms proposed in this dissertation has been significantly improved and the algorithms can efifectively resist the attacks from common image steganalysis algorithms and image tampering forensic algorithms.