Research On Steganographic Security Enhancement And Distribution Preserving Steganography

The continuous development of information and communication technology and the popularization of mobile Internet have made information transmission convenient.Information about military,politics,and business that are very sensitive to countries,enterprises,or individuals is transmitted or stored on the Internet.In order to protect these sensitive data,traditional methods usually encrypt sensitive data into meaning-less cipher text.However,the act of encryption exposes the existence of sensitive data,which is vulnerable to attack.The steganography can be used as a supplement to the encryption technique.It can hide both the secret information and the behavior of trans-mitting or storing secret information.However,with the rapid development of high-dimensional feature steganalysis and deep learning based steganalysis,steganography is facing new challenges.That is to say,the traditional steganographic methods are easier to be detected.Therefore,it is of great significance to study steganographic security enhancement methods and strate-gies.There are many methods to improve the security of steganography,including the following three directions:First define steganographic distortion more precisely to re-duce the damage of steganography modification to the correlation of neighbor elements.The second is to design non-additive steganography to keep the neighborhood relevance of the elements as much as possible.The third is designing distribution preserving steganography.The main work and innovations of this dissertation are as follows:1.Microscale Steganographic Distortion Model for Image SteganographyObserving the distribution of modifications of the existing image adaptive steganog-raphy algorithm,we found that there are still many modifications in the smooth area,indicating that the current definition of steganography distortion is not fine enough.With the help of image enhancement technology,this dissertation proposes a microscale steganographic distortion model.The steganographic distortion function is defined on the enhanced image and then assigned to the original image.In this way,the stegano-graphic modification can be concentrated in the texture area.Besides,with the help of DCT domain filtering,the microscale distortion model is applied to JPEG image steganography,and the efficiency of the algorithm is theoretically proved.Experimen-tal results show that the seed distortion function applying the proposed microscale dis-tortion model,has stronger security.2.Reversible Steganography Algorithm for JPEG image based on Multi-level Distortion Recursive CodingBased on the characteristics of JPEG images,a coefficient selection strategy is pro?posed.According to the influence of the modification on the spatial domain,the modi-fication distortion of the coefficient is defined.Then a DCT block selection is proposed by solving the optimization problem of minimizing distortion.Finally,through multi-level distortion recursive coding,information embedding,extraction and cover signal recovery are completed.Experiments show that the method proposed in this disserta-tion can effectively maintain the visual quality of the image,reduce bit rate expansion,and improve the security against steganalysis.3.Distortion Definition for Audio Adaptive Steganography and its Non-additive ExtensionsWith the development of mobile communication applications and the large-scale transmission of audio on the network,the study of audio adaptive steganography is of great significance to information security.According to the characteristic that the low amplitude region of audio is easy to be modeled,the principle of large amplitude priority is proposed,and its rationality is verified through experiments.Since the state-of-the-art steganalysis methods are designed on derivative of audio,we propose the steganographic distortion based on derivative filter residuals,guided by the principle of large amplitude priority.Considering that the steganography modification affects each other,a non-additive scheme for audio steganography is designed.Experimental results show that the algorithm's anti-detection ability has been significantly improved compared with existing steganographic algorithms.At the same time,the non-additive scheme further improves the steganographic security.4.Distribution-preserving Steganographic Algorithm based on Deep Gener-ative ModelDeep learning has made the steganalysis more effective,and also brought oppor-tunities to steganography.The generated data based on the deep generative model is widely used on the network and meet the requirement of being a digital cover.Besides,the generative model provides us with an explicit probability distribution of the gener-ated data or a sampler with the same distribution.In this paper,based on the explicit generation model as well as arithmetic coding,a distribution preserving steganography algorithm based on decompression is constructed,and its security is proved from the perspective of information theory.Based on the reversible implicit generative model and rejection sampling,the distribution preserving steganographic algorithm based on sampling is designed.In practice,this dissertation builds two types of steganographic systems based on the speech synthesis models WaveNet and Waveglow,respectively.The experiments show that the steganalysis methods cannot distinguish whether the audio is good or evil,indicating that the distribution of cover is well preserved.