Research On Key Techniques Of Image Steganalysis For GAN-based Steganography

Steganography hides information in the carrier of digital objects for safe transmission,so as to open up a concealed and secure communication channel over public networks,and steganalysis restricts the malicious use of steganography by criminals.The two types of technologies have been in a state of alternating rise for a long time.Recently,the emergence of steganography based on deep learning has broken the relative balance between steganography and steganalysis.The combination of generative adversarial network(GAN)technology and steganography has exerted the powerful self-learning and continuous optimization capabilities,and significantly improved the performance of the steganography algorithm.On the one hand,it greatly reduces the dependence on manual design.On the other hand,the security of the GAN-based steganography is rapidly improved,that is,the detection accuracy of mainstream steganalysis models is significantly reduced.So far,there is no effective detection method for GAN-based steganography algorithm.Therefore,it is of great significance to master the method of detecting GAN-based steganographic samples.To solve this problem,the thesis conducts research and improvement from the perspective of security sources of steganographic algorithms based on generative adversarial networks.The main work and contributions are as follows:1.This thesis studies the security of GAN-based steganography algorithm,and compares it with adversarial samples from three perspectives:network architecture analysis,decision plane analysis,network effect and target analysis.It is pointed out that its security builds on the non-robust characteristic similar to adversarial samples,which gives a feasible research direction for the improvement of detection accuracy.Then,the security of the GAN-based steganalysis algorithm is evaluated by using two types of detection methods:a steganalysis algorithm based on handcrafted features and a steganalysis algorithm based on deep learning.The experiments prove that the GAN-based steganography has stronger anti-detection ability than the cutting-edge content adaptive steganographic algorithm.Finally,the actual embedding change position differences are analyzed between GAN-based steganography and content adaptive steganography.2.Aiming at the discovery that the security of the GAN-based steganography comes from the "non-robust features",a novel " Hybrid Training " strategy is proposed for the first time to break the security and concealment based on the generative confrontation network.This "Hybrid Training" strategy can be regarded as a diversity-based training strategy,which guides the network to fully learn steganographic features by mixing GAN-based steganographic samples and other types of steganographic samples during the training process.According to the two variables of different mixing ratios and different mixing steganography types,an algorithm for searching the optimal mixing method is designed.And the influence of different mixing types and mixing ratios on the detection effect is analyzed through extensive experiments.Experiments prove that the proposed method can improve the detection accuracy by 7.96%compared with a single training method in the challenging detection of steganographic images.Besides,this inference is verified through the two visualization methods of Grad-CAM and T-SINE.In addition,in the actual situation of parameter mismatch,the proposed method can better detect other steganography algorithms including adaptive steganography,so it is more practical in image detection.3.Inspired by the GAN-based steganography algorithm,this thesis shifts the training focus of the generative adversarial network from the generator to the discriminator.By distinguishing the steganographic image and the cover image generated by the generator,mutual supervision and continuous self-optimization make the discrimination as accurate as possible,thus establishing a powerful discriminant model.In order to solve the problem caused by the inconsistent parameters between training and generation process,the training process is divided into two stages,so as to close the realistic process from the perspective of samples.Besides,the design of the loss function of the generator and the discriminator is further optimized.On the basis of the original GAN-based steganalysis algorithm,the image distance loss and the feature distance loss are respectively introduced as the direction and goal of the training of the generative adversarial network.From the experimental results,the detection accuracy of the two improved methods based on GAN-based steganalysis has been significantly improved.