Research On Digital Steganography Based On Wav Audi

Digital steganography is an important means to ensure the security of data transmission.Audio is a common digital medium and digital steganography based on audio has been studied in depth.However,there are still some problems with audio steganography technology at present.On the one hand,for the audio steganography by cover modifications,Researchers are committed to studying how to improve the undetectability of algorithms,but steganography capacity is easily constrained,which means that both cannot be improved simultaneously.On the other hand,the current generative audio steganography still relies on an existing or generated cover to carry secret messages,which makes it unable to fundamentally resist the detection of the existing steganalysis tools.In order to solve the above problems,this thesis proposes corresponding audio steganography algorithms.The main work is as follows:（1）An adaptive universal audio steganography algorithm based on genetic algorithm is proposed.Firstly,this work changes the grouping method based on time in the traditional modified steganography and proposes a dynamic grouping method based on sampled values.Secondly,genetic algorithm is used to determine the optimal embedding position in each group through the iterations of population.KL divergence is used to calculate the fitness value of the genetic algorithm.In addition,this work proposes a dynamic embedding method and the modified values of each embedding point are jointly influenced by the sampling value,secret message sequence,and grouping rules.Ultimately,modifying one sampling value can embed multiple secret messages.At last,experimental comparisons and analyses were conducted on embedding algorithms with different parameter settings.Under the default parameter settings,the steganography capacity is 3.2kbit/s.The average PESQ of stego-audio reaches 4.20 or above,and the SNR reaches 120db or more,which proves that stego-audios have good speech quality.Its KL divergence value reaches 10^-10 levels,indicating that the algorithm has good nondetectability.And when the limit capacity is 19.5kbit/s,the KL divergence value reaches 10^-16 levels,indicating that this method improves both steganography capacity and nondetectability.This method can embed the secret messages of any length by adjusting the audio duration,sampling rate,and various parameter values of this embedding algorithm,which improvies the universality of this steganography algorithm.（2）A coverless audio steganography model based on Generative adversarial networks is proposed.It consists of two parts:a generation module and a reconstruction module.On the one hand,the generation module,based on the speech synthesis model Wave GAN,is employed to directly generate stego-audios driven by secret audios.This work modifies the input of the generator in Wave GAN to secret audio instead of random noises,and adds a post-processing layer to enhance the speech quality of the generated stego-audios.On the other hand,this work designed a reconstruction module to reconstruct the secret audios.It is understood that this is the first audio steganography method to directly generate stego-audios.Experimental results show that it is difficult for the current steganalysis methods to detect the existence of secret messages,because this method does not perform any modifications to existing or generated cover.The generated stego-audios hava high speech quality,and their average MOS is 4.13.Besides,the steganography capacity can be measured from two perspectives:from the perspective of size,the steganography capacity can reach 50%;from a semantic perspective,22-37 bits can be hidden in a two-second stego-audio.Subsequently,it is proved by the spectrums in different forms that the reconstruction module can reconstruct the secret audio successfully on hearing.STOI and the speech recognition detections certify that the reconstruction module can ensure the complete semantic transmission of secret audios.Finally,after adding random noise to the stego-audios,the reconstruction module can still reconstruct the secret audio successfully to some extent,which states that the proposed method has good robustness.