Research On Constructive Information Hiding Method Based On Deep Learning

Information hiding is a technology that hides secret information into a multimedia carrier.By using this technique,the sender and the receiver can communicate without arousing suspicion of third parties.The traditional information hiding method is to modify the multimedia carrier directly according to the secret information,and the modification of the carrier will inevitably lead to the distortion of the carrier.With the rapid development of steganalysis technology,steganalyzers can capture increasingly subtle carrier distortions,which directly threaten the security of traditional information hiding methods.In order to fundamentally ensure the security of information hiding techniques,researchers proposed generative information hiding technology.The generative information hiding technique is a direct construction of steganographic carriers under the guidance of secret information.There is no trace of modification in the steganographic carrier,so it can effectively resist steganalysis.However,current generative steganography methods have very limited hiding capacity in order to ensure the quality of steganographic carriers and the extraction rate of secret information.So,this paper proposes two deep learning-based generative information hiding schemes,and the detailed researches are as follows:(1)A secret-to-image reversible transformation(S2IRT)scheme based on flow-based model and position coding for constructive steganography is proposed.Guided by a given secret message,we encode a position index and then construct a latent vector by it.Fanally,we map the latent vector to a generated image by the Glow model,so that the secret message is finally transformed to the generated image.The flow-based model enables a bijective-mapping between latent space with multivariate Gaussian distribution and image space with a complex distribution,which results in higher hidden capacity and extraction accuracy.Compared with existing methods,the proposed steganographic approache can achieve high hiding capacity(up to 4 bpp)and accurate information extraction(almost 100% accuracy rate)simultaneously,while maintaining desirable anti-detectability and imperceptibility.(2)A secret-to-image reversible transformation scheme based on flow-based model and improved position encoding(SE-S2IRT)for constructive steganography is proposed.SES2 IRT improves S2 IRT by using separate encoding to enhance robustness.The experiments demonstrate SE-S2 IRT has better robustness than S2 IRT.(3)A constructive behavior steganography scheme based on LSTM is proposed.This scheme encodes the secret message into the behaviors performed by the sender in the future.First,the trained LSTM model is used to predict the next behavioral state.Then high probability behaviors are screened to construct a behavior tree,the tree contains possible future behaviors.Finally,the behavioral state sequence is selected guided by secret message.This method has higher generalizability than single carrier-based steganography methods.