Research On Image Additive Steganography Based On Adversarial Optimizatio

Image steganography is a significant research field in information security,as it serves as a primary branch of multimedia steganography and finds extensive application in privacy protection and covert communication.This technique achieves the transmission of secret messages by modifying the redundant information of images while minimizing the distortion.However,the emergence of deep steganalysis has revealed several issues in existing image steganography.The widely used deep hiding,for instance,faces problems such as low security and poor extraction quality.Similarly,heuristic adaptive steganography encounter problems such as complicated manual design,embedding cost incompatibility across various domains,and the difficulty of deep distortion training.To address these problems,this paper focuses on the training process of the deep hiding algorithm and explores the application of deep learning in heuristic distortion design.It proposes targeted solutions to enhance the visual quality and communication concealment of the covers.The research contents of this paper include the following:(1)This thesis proposes the A-D-UNet algorithm based on cross-domain adversarial adaptation to address the issue of poor image quality in existing deep hiding.The algorithm includes a super-resolution covers enhancement scheme that embeds secret information into image content with scale invariance.The self-attention mechanism is introduced into the encoder to guide the model to focus on content features and enhance stegos generation quality.Additionally,to address the inadequate concealment of steganography,A-D-UNet incorporates domain-adaptive loss into the encoding network for adversarial training to guide the generation of covers and reduce the cross-domain difference between covers and stegos.Experimental results demonstrate that A-D-UNet improves security while ensuring image quality when compared to similar algorithms.(2)This thesis addresses the issues of decreased steganographic security and statistical anomalies caused by the aggregation of additive distortion towards complex texture boundaries when using existing additive heuristics steganography under deep analysis features.From the perspective of the relationship between embedding distortion cost and texture sparsity,the paper proposes a maximum-minimum entropy clustering optimization strategy using the Canny operator to divide textures and Gaussian blur to scale contours.In combination with Auto ML technique adversarial enhancement of deep steganalysis features,a universal additive steganographic algorithm called Canny Gauss is proposed.Experimental results demonstrate that Canny Gauss exhibits higher invisibility and embedding cost stability in spatial,JPEG,and two types of side-informed embedding domains.(3)This thesis introduces Entropy Filter(EF),a deep information entropy adversarial optimization-based additive steganography that addresses the problem of inadequate steganography security caused by abnormal expression of embedding costs from multiple simulation embedding vanishing gradient and excessive resistance to the special steganalysis features.The EF algorithm employs the bi-level optimization strategy to calculate the weighted expectation of the average distortion payload,replacing the binary search solution to complete the backpropagation and avoiding ineffective optimization.Additionally,the algorithm utilizes global minimum entropy and local maximum entropy inner adversarial losses to improve steganography distortion design and replace the mainstream outer adversarial learning schemes.Experimental results demonstrate that EF achieves a security performance comparable to manually designed additive distortion without introducing additional steganalysis features.