Research On Multimedia Steganography And Steganalysis Based On Deep Learning

Information security in cyberspace is crucial to national security and social stability.With the acceleration of information construction,the problem of data security in cyberspace is increasingly exposed.Massive multimedia data are vulnerable to security threats such as eavesdropping,tampering and disclosure.Information security and privacy protection have attracted more and more attention.Information hiding technology in cyberspace has always been a strategic research field that countries around the world pay great attention.The development of network technology,artificial intelligence and big data provides new ideas and data foundations for information hiding.Just like the red and blue parts in cyberspace,steganography and steganalysis have always been in a situation of mutual promotion and development in the confrontation game.In recent years,steganalysis technology based on deep learning has made great progress,and deep learning networks can build better steganalysis features,thereby greatly improving the detection accuracy.The advancement of deep steganalysis threatens the security of traditional steganography,the security of steganography needs to be improved.A secure steganography algorithm can be generated without any human involvement by adversarial training.Using adversarial technology can effectively improve the security of steganography by taking advantage of the vulnerability of deep model.This paper focuses on the following researches on the confrontation problem of multimedia steganography and steganalysis based on deep learning.In deep learning steganalysis,aiming at problem that the current deep steganalysis network could not converge smoothly,the multi-dimensional convolution network structure is used to extract richer steganalysis features.In order to reduce the loss of steganographic feature information between network layers in steganalysis,the maximum pooling layer is reduced when designing the network structure,and the normalization layer is used to improve the efficiency of model training.In addition,in view of the problem that the steganalysis performance of the proposed deep model degrades at low embedding rate,a transfer learning strategy is introduced,and the small noise residual features at low embedding rate are learned through progressive training.The effectiveness of the proposed deep steganalysis method is verified by experimental analysis.In the steganalysis under high embedding rate,the detection performance of the proposed method is better than the classical rich model and the typical deep steganalysis network Ye.Net.The detection performance is better than that of classical rich models at low embedding rates,and is comparable to that of typical deep steganalysis networks.In the model training experiment based on transfer learning,the model can gradually converge,indicating that the model has the detection ability of small steganographic noise,and the detection ability of steganalysis is significantly improved.In the aspect of deep steganography,a speech steganography model based on generative adversarial network is proposed.The improvement of traditional steganography methods relies on the defects of manual design,and introduces the idea of adversarial game.By simulating the confrontation process in information hiding,a generative steganography scheme based on generative adversarial network is designed.All parameters of the steganography algorithm are obtained through game learning.First,the audio is converted into a spectrogram using fast Fourier transform,and then the information is embedded and extracted through the encoding and decoding network,and finally the steganographic audio samples are obtained through inverse transformation.A steganalyzer is added to the model training for simultaneous training,and the truncation gradient is used to optimize the steganographic ability of the encoder.In the process of model training,the encoding network and the decoding network have the function of steganography by designing their respective loss functions.The steganalyzer acts as the supervisor of the model training,and the loss function is the cross entropy of the carrier and the carrier.Finally,relevant experiments are carried out to compare with similar algorithms.The steganographic method based on generative adversarial network has a good steganographic effect.Due to the process of adversarial game,the proposed steganography method has a certain ability to resist deep steganalysis detection.Aiming at the problem that the security of adaptive steganography is difficult to guarantee due to the increasingly powerful feature abstraction ability of deep steganalysis,a security improvement method of adaptive steganography based on adversarial samples is proposed.Using the linear characteristics of deep steganalysis in high-dimensional space,we carefully design imperceptible tiny anti-noise to mislead the deep steganalysis model and protect the features of adaptive steganography.Both adversarial noise and steganography are minor changes to the original image,both of which are general,and it is feasible to use adversarial example techniques to mislead deep steganalyzers.To improve the problem of negative optimization of steganography ability caused by anti-noise embedding and steganographic noise block processing in the previous work,a more secure steganographic framework is proposed.Accurate extraction of adaptive steganographic samples.In the process of adversarial noise generation,the adversarial attack method with less disturbance is adopted,which greatly improves the security of steganography.In the aspect of generative steganography,an improved generative steganography method based on adversarial samples is proposed in view of the inherent defect that generative steganography reduces security due to the increase of steganographic capacity.Different from the traditional hand-designed steganography,the generative steganography method obtains the specific parameters of the steganography algorithm by constructing a deep network and then training.Compared with the traditional steganography method,the steganographic capacity is greatly improved,although the capacity is greatly Improved,but less secure against steganalysis.In view of the above problems,the generated steganographic samples are further strengthened by using the adversarial sample technology.The improved scheme first goes through the steganographic encoder,and then adopts the method of anti-noise embedding.In terms of perturbation selection,both confrontation and concealment are taken into account.Different adversarial sample attacks are used for experiments under two different datasets,and the final experimental results verify the effectiveness of the proposed method.