Non-embedding image zero-distortion steganography method and system

Abstract

The invention belongs to the technical field of image steganography, and particularly relates to an embedding-free image zero-distortion steganography method and system. The method comprises the following steps: S1, carrying out tensorization processing on secret information; s2, generating a steganographic image based on the secret information after the tensorization processing; and S3, re-extracting the secret information from the generated steganographic image. The method has the characteristics that the steganographic image generation quality can be enhanced while the steganographic analysis attack resistance can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image steganography, and in particular relates to a method and system for image zero-distortion steganography without embedding. Background Art

[0002] Information hiding technology ensures the confidentiality and security of communications by concealing secret information within carrier data (such as text, audio, and images). Its core requirements include imperceptibility, anti-interference capabilities, and information capacity. Image steganography, a key branch of information hiding, exploits the redundancy of digital images (such as pixel spatial redundancy, frequency domain coefficient redundancy, or semantic redundancy) to integrate secret information into the carrier image, thereby avoiding suspicion from third parties.

[0003] Traditional image steganography techniques are primarily categorized as embedded steganography and generative steganography. Embedded steganography embeds secret information by modifying redundant regions of the carrier image (such as frequency domain transform coefficients, high-texture regions, or edge regions). For example, frequency domain transform methods modify frequency domain coefficients based on discrete cosine transform (DCT) or discrete wavelet transform (DWT) and optimize distortion functions to minimize the impact of the embedding operation on visual quality. Spatial domain adaptive methods dynamically adjust the embedding strength based on image texture complexity to balance stealth and capacity. Deep learning-driven methods utilize generative adversarial networks (GANs) or convolutional neural networks (CNNs) to learn optimal embedding strategies to minimize distortion of the carrier image. However, these methods, because they require direct modification of the carrier image, can easily lead to anomalies in statistical characteristics (such as frequency domain energy distribution and pixel correlation), making them detectable by steganalysis tools (such as StegExpose and XuNet). Furthermore, high-capacity embedding operations can significantly degrade image quality metrics such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), making it difficult to achieve a balance between stealth and fidelity.

[0004] To circumvent these issues, generative steganography abandons modification of the base image and directly utilizes generative models (such as GANs and diffusion models) to synthesize stego images containing secret information. Typical methods include latent vector mapping (encoding the secret information into the latent vector of the generative model to drive image synthesis) and feature fusion (fusing the secret information with the structural or texture features of the base image to generate the stego image). While these methods can avoid detection by traditional steganalysis, they still face the following technical bottlenecks: First, the instability of the generation algorithm can easily lead to structural distortion or texture distortion in the stego image, which can be detected by detection methods based on perceptual similarity (LPIPS) or image quality assessment (such as PSNR and SSIM). Second, existing methods lack sufficient constraints on the structural features of the image, resulting in a monotonous stego image style and difficulty adapting to the diverse needs of complex scenes. Third, some techniques (such as the IDEAS method) rely on the secret information to directly drive the generation process and lack deep integration of the structural features of the reference image. This makes the generated results susceptible to noise and attacks, and has weak anti-disturbance capabilities.

[0005] Therefore, it is very important to design a non-embedded image zero-distortion steganography method and system that can improve the ability to resist steganalysis attacks while enhancing the quality of steganalysis image generation. Summary of the Invention

[0006] The present invention aims to overcome the problem of instability of generative image steganography technology in the prior art during image generation, and provides a non-embedded image zero-distortion steganography method and system that can improve the ability to resist steganalysis attacks while enhancing the quality of stego-image generation.

[0007] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0008] The zero-distortion image steganography method without embedding includes the following steps:

[0009] S1, performs tensor processing on secret information;

[0010] S2, generates a stego-image based on the secret information after tensorization;

[0011] S3, re-extracts the secret information from the generated stego-image.

[0012] Preferably, step S1 includes the following steps:

[0013] Step S11: Divide the secret information M into segments of length ε bits, each segment is denoted as m; map the integer secret message m to the floating point number f, that is:

[0014]

[0015] Where ε is a custom constant;

[0016] Step S12: introducing a random interval into the floating point number f to enhance the diversity of the generated image, namely:

[0017]

[0018] Among them, z is the floating point value corresponding to the secret message m; the range of the random interval is controlled by the hyperparameter η, and the total size is η×2 2-ε , the range is [-η×2 1-ε ,η×2 1-ε ]; rand represents a random number;

[0019] Step S13: Concatenate the randomized information z in the set order to form a secret information tensor Z M .

[0020] Preferably, step S2 includes the following steps:

[0021] S21, train the encoder E to obtain the structure information tensor S and texture information tensor T corresponding to the input image I, that is:

[0022] S,T=E(I);

[0023] S22, design a multi-scale attention feature fusion MSAFF method; the MSAFF method includes a generator G stru and fusion device E fusion ;

[0024] The generator G stru , used to transform the secret information tensor Z M Encoded into the preliminary structure tensor S related to the image structure information tensor M middle;

[0025] The fusion device E fusion , used to combine the image structure information tensor S obtained in step S21 and the generator G stru Generated preliminary structure tensor S M Fusion is performed to obtain enhanced structural information tensor

[0026] S23, introduce a stego image generator G based on a GAN network, and generate different images by inputting different combinations of structure information tensors and texture information tensors T; the combination of structure information tensors and texture information tensors T includes the original structure information tensor S and texture information tensor T, the preliminary structure tensor S M With texture information tensor T, enhanced structure information tensor With the texture information tensor T

[0027] S24, introduce the discriminator D co and a discriminator D; the discriminator D co At the same time, the texture loss between the image generated in step S23 and the input image I is calculated; the discriminator D is used to optimize the steganographic image generator G to ensure the authenticity and naturalness of the generated image.

[0028] As a preference, in step S21, during encoder training, a discriminator D is introduced. dist Encourage the texture information tensor T to follow a specific distribution, then the loss function L during training E,dist is defined as:

[0029] L E,dist =softplus(-D dist (T));

[0030] The softplus function is defined as: softplus(x)=log(1+e x ).

[0031] As a preference, the fusion device E fusion The specific process includes the following:

[0032] S221 uses the multi-scale attention feature fusion MSAFF method to calculate the fused weight value to enhance the expression ability of structured information. The specific formula is expressed as:

[0033]

[0034] Among them, σ represents the sigmoid activation function; MLP represents the multi-layer perceptron; AvgPool represents the average pooling; MaxPool represents the maximum pooling, Indicates the initial feature integration. In order to simplify the calculation, the element-level summation is selected as the initial integration, and the calculation result is Denoted as symbol α; then the enhanced structural information for:

[0035]

[0036] in, represents element-wise product; Represents the final structured information of the stego-image; Represents the real space, C represents the number of channels, H represents the number of pixels in the height of the input image, and W represents the number of pixels in the width of the input image; the weight value after fusion That is, α is a real number between 0 and 1, which is used to control the network between S and S M Parameter selection ratio between ;

[0037] S222, introduce a discriminator D dist_struc , so that the fused structured information The distribution of is approximately equal to S, then the corresponding loss function is defined as:

[0038]

[0039] final, The feature distribution of will tend to the feature distribution of the real image.

[0040] Preferably, step S23 further includes the following steps:

[0041] S231, based on the combination of different input structure information tensors and texture information tensors T, three different stego images are obtained, namely:

[0042] I1=G(S,T);

[0043]

[0044] Among them, I2 and I3 are steganographic images containing secret information, and the corresponding loss function L G,rec as follows:

[0045] L G,rec =||I1-I|| 1 ;

[0046] Among them, ||·|| 1 Represents L1 loss, which is the absolute loss function.

[0047] Preferably, step S24 includes the following steps:

[0048] S241, using the discriminator D co Compute the texture loss between I2 and I; discriminator D co Cut out an image block from each of the stego image I2 and the input image I, encode and connect the cut image blocks, and use the network to calculate the loss L during training. G,texture :

[0049] L G,texture =D co (patch(I2),patch(I));

[0050] Among them, the function Patch(·) represents the cropping function;

[0051] S242, introduce the discriminator D to optimize the generator G and the overall loss function L G,real as follows:

[0052] LG,real =D(I1,I)+D(I2,I)+D(I3,I).

[0053] Preferably, step S3 includes the following steps:

[0054] S31, the extractor E extracts the corresponding structural information from the stego-image I2 or I3 Right now:

[0055]

[0056] The loss function L when training the extractor E E,struc for: The extractor E shares parameters with the encoder E;

[0057] S32, Extractor E fusion,ex Further from structured information Extract the tensor corresponding to the secret information M from Right now:

[0058]

[0059] Training the extractor E fusion,ex The loss function L when E,fusion for:

[0060] S33, another extractor E x from Extracting secret information Right now:

[0061]

[0062] Training the extractor E x The loss function L when Ex for:

[0063] The present invention also provides an embedding-free image zero-distortion steganography system, comprising:

[0064] A secret information processing module, used for performing tensor processing on secret information;

[0065] A stego image generation module, used to generate a stego image based on the secret information after tensorization processing;

[0066] The secret information extraction module is used to re-extract the secret information from the generated stego image.

[0067] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention improves the visual quality and confidentiality of the generated stego image by encoding and fusing the structural information and texture information of the carrier image; (2) The present invention deeply fuses the secret information with the structural representation of the reference image through the Multi-Scale Attentional Feature Fusion (MSAFF) algorithm to generate structured information features with enhanced stability, thereby improving the stability and accuracy of the steganographic system in the process of information embedding and extraction; (3) The present invention improves the concealment and anti-detection ability of the stego image by optimizing the steganographic method. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flow chart of the non-embedded image zero-distortion steganography method of the present invention;

[0069] Figure 2 A schematic diagram of the images obtained by various generative steganography methods. DETAILED DESCRIPTION

[0070] To more clearly illustrate the embodiments of the present invention, specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings and other embodiments can be obtained based on these drawings without inventive efforts.

[0071] like Figure 1 The figure shows a zero-distortion steganography method without embedding in the present invention. This method does not rely on directly modifying the reference image to embed secret information, but generates a new stego image based on the secret information, fundamentally avoiding the detection of the steganalysis algorithm. In addition, to address the image distortion and distortion problems that may be caused by the generation algorithm, the present invention further proposes the MSAFF method. This method effectively improves the structural stability and authenticity of the generated secret image by fusing the secret information with the structured information of the real image, and improves the accuracy of secret information extraction. The specific steps are as follows: Figure 1 As shown:

[0072] Step 1: Perform tensor quantization on the secret information, as follows:

[0073] Step 1.1: Split the secret message M into segments of length ε bits, each segment is denoted as m. Map the integer secret message m to the floating point number f, that is: Where ε is a custom constant.

[0074] Step 1.2: Introduce a random interval for the floating-point number f to enhance the diversity of the generated image, where z is the final floating-point value corresponding to the secret message m, that is: The range of the random interval is controlled by the hyperparameter η, and the total size is η×2 2-ε , the range is [-η×2 1-ε ,η×2 1-ε ].

[0075] Step 1.3: Concatenate the randomized information z in sequence to form the secret information tensor Z M In particular, when the extractor converts the tensor Z M When restored to secret information m, in It represents rounding up. By sequentially concatenating the restored secret information m, the complete secret information M can be reconstructed.

[0076] Step 2: Generate a stego image based on the secret information after tensorization, as follows:

[0077] Step 2.1: Encoder E training: The encoder E is used to encode an image into corresponding structural information and texture information. It takes an image I as input, randomly samples from the dataset, and outputs the input structural information S and texture information T corresponding to the image, that is: S, T = E(I). In particular, during encoder training, the discriminator D dist Encourage the texture information T to follow a specific distribution, where the loss function during training can be defined as L E,dist =softplus(-D dist (T)), where the softplus function is defined as: softplus(x) = log(1 + e x ).

[0078] Design of Multi-Scale Attention Feature Fusion Algorithm (MSAFF): This algorithm generates structured information features with enhanced stability by deeply fusing secret information with the structural representation of the reference image. The MSAFF algorithm here consists of a generator G stru and fusion device E fusion Composition, respectively introduced:

[0079] Generator G stru The role of the secret message Z M Structural information S encoded into the image M To support the subsequent feature fusion process. First, the secret message M is transformed into the corresponding secret message tensor Z through the mapping rule described in step 1.3 M , Z M As the generator G stru Input, used to generate a preliminary structured tensor S representing the secret message M .

[0080] Step 2.2: Fusion Cage E fusion The role of is to combine the structure tensor S obtained in step 2.1 and the preliminary structure information S M The fusion is performed to provide support for the subsequent generation of steganographic images. The multi-scale attention feature fusion (MSAFF) method used in this invention fuses S and S M The weight of is used to enhance the expression ability of structured information. The specific fusion formula is: in Represents the initial feature integration. In order to simplify the calculation, the element-level summation is selected as the initial integration, and the symbol α is The calculation results of enhanced structured information for: in represents element-wise product, Represents the final structured information of the stego-image, It is worth noting that the weight value after fusion That is, α is a real number between 0 and 1, which is used to control the network between S and S M In addition, in order to make the final generated target image have more realistic semantic meaning in the real world, the discriminator D is introduced when training the fusion. dist_struc , so that the fused features The distribution of is approximately the same as S, and the loss function here is:

[0081] Step 2.3: Steganographic image generation: After obtaining the structured information mentioned in step 2.2 Finally, StyleGAN2 is used as the core architecture of the Generator to synthesize more realistic stego images. The stego image generator here takes structural information S and texture information T as input and outputs the corresponding image. Based on different input combinations of S and T, three different stego images can be obtained, namely: I1=G(S,T), Among them, I2 and I3 are steganographic images containing secret information. The corresponding loss function is as follows: L G,rec =||I1-I|| 1 , where ||·|| 1 represents L1 loss.

[0082] Step 2.4: Improvement of stego-image quality: Based on the StyleGAN2 mentioned in step 2.3, the discriminator D is introduced. co and D. Discriminator D co Specifically used to calculate the texture loss between I2 and I. It crops small image patches from I2 and I, encodes them, connects them together, and uses the network to calculate the loss during training: LG,texture =D co (patch(I2),patch(I)), where the function Patch(·) represents the cropping function. In addition, in order to ensure the authenticity and naturalness of the generated image, the discriminator D is introduced to optimize the generator G. The overall loss function is as follows: L G,real =D(I1,I)+D(I2,I)+D(I3,I), where D indicates that these discriminators have the same architecture.

[0083] Step 3: Secret information extraction: The embedding-free image zero-distortion steganography framework proposed in this invention involves multiple stages and multiple extractors in secret information extraction.

[0084] First, the extractor E extracts the corresponding structural information from the stego-image I2(I3) Right now: The loss function when training the extractor E is: Then, the extractor E fusion,ex Further from structured information Extract the tensor corresponding to the secret information Right now: Training the extractor E fusion,ex The loss function when is: Finally, another extractor E x from Extracting secret information Right now: Training the extractor E x The loss function when is:

[0085] The ZeroDS proposed in this paper divides the overall loss L into two main parts, namely the loss L for stego-image generation and the loss L for stego-image generation. G and the loss L for secret message extraction E , which can be expressed as follows:

[0086] L=L G +L E

[0087] L G The specific expression is as follows:

[0088] L G =L G,rec +L G,texture +2×L G,real

[0089] You can use L G,real The coefficient 2 before L is used to improve the quality of the generated image. E The specific expression is as follows:

[0090] L E =L E,dist +L E,struc +L E,fusion +γL Ex

[0091] Here, γ is a hyperparameter used to balance the trade-off between secret message extraction accuracy and image generation quality, and is set to 10 in the experiments.

[0092] The remarkable technical effects of the present invention are further illustrated below in conjunction with experiments.

[0093] The experiments were conducted on a high-performance server running the Ubuntu 18.04.5 operating system. The hardware configuration included an Intel Xeon Silver 4210R processor (40 cores, 2.40GHz) and four NVIDIA RTX 3090 graphics cards with a total memory capacity of 96GB. The experimental datasets used were FFHQ and LSUN. FFHQ contains 70,000 face images, while LSUN uses the training set of the Bedroom and Church subsets. All images were resized to 256×256 resolution to accommodate the model input requirements. In the experiments, the parameter ε was set to 1, η was set to 0, and γ was set to 10.

[0094] The performance evaluation of this method considers three metrics: generation quality, concealment security, and information fidelity. Experimental comparisons were conducted using the FFHQ and LSUN datasets, respectively, with mainstream generative steganography methods. Generation quality evaluation quantitatively analyzes the visual quality and structural consistency of the image; concealment security analysis quantitatively verifies the concealment and authenticity of the stego-image; and information fidelity analysis quantitatively measures the accuracy of recovering the secret information from the stego-image.

[0095] The FID (Fréchet Inception Distance) value is used to analyze the quality of steganalytic image generation. The lower the FID value, the closer the distribution of the generated image is to the real image. The details are shown in Table 1 below:

[0096] Table 1 Evaluation data of FID values of steganographic images under different datasets

[0097]

[0098]

[0099] DCGAN-Steg is an embedding-free image steganography method based on deep convolutional generative adversarial networks. This method maps secret information into a noise vector and uses a trained generator to generate a carrier image, eliminating the need to modify the image content. The trained extractor network can effectively recover the secret information contained in the image.

[0100] SAGAN-Steg is an embedding-free image steganography method based on the attention mechanism GAN. By improving the generator, discriminator and extractor structures, it effectively improves image quality, steganographic capacity and information recovery accuracy.

[0101] SSteGAN is a secure steganography strategy based on generative adversarial networks. It uses a generator and two discriminators to collaboratively generate high-quality images suitable for steganography, and has achieved significant results in improving image quality, training stability, and convergence speed.

[0102] WGAN-Steg is a novel, non-modification image steganography method based on the Wasserstein GAN gradient penalty (WGAN-GP). This method achieves steganography by synchronously training a generator (which generates the vector and stego-image) and an extractor (which extracts the secret information), without modifying the image.

[0103] IDEAS stands for Image Decoupled Autoencoder Steganography, an embedding-free steganography technique that converts secret messages into synthetic images for hiding, which is resistant to traditional steganalysis.

[0104] from Figure 2 It can be seen from the images obtained by each generative steganography method and the FID values of the images obtained by each generative steganography method in Table 1 that the FID value of the method proposed in the present invention is the lowest among all models, showing the best generated image quality.

[0105] Three well-known steganalysis tools were used to analyze the hidden security of steganalysis images, including three network architecture types: StegExpose, XuNet, and YeNet. These tools cover traditional steganalysis algorithms and deep learning-based detection methods, and can provide a comprehensive security assessment of steganalysis images. To quantify the detection performance, we performed a receiver operating characteristic (ROC) curve analysis on the generated steganalysis images and calculated the area under the curve (AUC) of the ROC curve. The closer the AUC value is to 0.5, the more difficult the steganalysis image is to detect and the higher the security. The experimental results are shown in Table 2, the AUC values under the ROC curve for each generative steganography method:

[0106] Table 2 AUC values under the ROC curve of various generative steganography methods

[0107]

[0108] As can be seen from the AUC values under the ROC curves for each generative steganography method in Table 2, all generative models have AUC values close to 0.5, indicating that the steganalysis tool's judgment of whether the model-generated image is a stego image is close to random guessing. This shows that these steganography methods are effective in resisting steganalysis attacks.

[0109] The stego-image distortion detection analysis verifies the fidelity of the generated image by evaluating the degree of deviation in visual quality and structural consistency between the stego-image and its target image. This experiment uses three common image quality evaluation metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Perceptual Image Similarity (LPIPS). The experimental results for each metric are shown in Table 3.

[0110] Table 3 Data table of distortion detection results of generated images of various generative steganography methods

[0111]

[0112]

[0113] Table 3 shows the distortion detection results for the generated images of various generative steganography methods. The proposed method outperforms other models in terms of PSNR, SSIM, and LPIPS. In Table 3, the symbol ↑ indicates that a higher metric is preferred, while the symbol ↓ indicates that a lower metric is preferred.

[0114] The fidelity analysis of stego-image information can reflect the robustness of the generative steganography method by quantitatively measuring the accuracy of the complete restoration of the secret information embedded in the stego-image. The experimental results are shown in Table 4, which shows the extraction accuracy of the images generated by each generative steganography method:

[0115] Table 4. Data on the extraction accuracy of images generated by various generative steganography methods

[0116]

[0117] In Table 4, Bedroom represents a dataset describing bedroom images, Church is a dataset representing churches, and FFHQ is a well-known dataset.

[0118] As can be seen from the extraction accuracy of the images generated by each generative steganography method in Table 4, the proposed method performs comparable to the IDEAS method in terms of extraction accuracy. In general, the extraction accuracy of these two methods is significantly better than that of the other comparison methods.

[0119] This paper proposes a zero-distortions teganography framework without embedding (ZeroDS) to address the image distortion, insufficient diversity, and limited ability to resist steganalysis that plague existing generative steganography methods. By encoding the structural and fine-grained texture information of a reference image and integrating the secret information with the structural information using a multi-scale attention feature fusion (MSAFF) algorithm, ZeroDS effectively enhances the stability of structural features. Furthermore, by randomly sampling texture information to generate style representations, the diversity of the stegoimages is significantly improved. Experimental results demonstrate that the ZeroDS framework demonstrates competitive performance in terms of steganalysis immunity, stegocapacity, and secret information extraction accuracy, while also significantly improving the visual quality and authenticity of the generated images.

[0120] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, based on the ideas provided by the present invention, there may be changes in the specific implementation methods, and these changes should also be considered as the scope of protection of the present invention.

Claims

1. The zero-distortion image steganography method without embedding is characterized by: The steps include: S1, performs tensor processing on secret information; S2, generates a stego-image based on the secret information after tensorization; S3, re-extracts the secret information from the generated stego-image.

2. The non-embedding image zero-distortion steganography method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Divide the secret information M into segments of length ε bits, each segment is denoted as m; map the integer secret message m to the floating point number f, that is: Where ε is a custom constant; Step S12: introducing a random interval into the floating point number f to enhance the diversity of the generated image, namely: Among them, z is the floating point value corresponding to the secret message m; the range of the random interval is controlled by the hyperparameter η, and the total size is η×2 2-ε , the range is [-η×2 1-ε ,η×2 1-ε ]; rand represents a random number; Step S13: Concatenate the randomized information z in the set order to form a secret information tensor Z M .

3. The non-embedding image zero-distortion steganography method according to claim 2, characterized in that: Step S2 includes the following steps: S21, train the encoder E to obtain the structure information tensor S and texture information tensor T corresponding to the input image I, that is: S,T=E(I); S22, design a multi-scale attention feature fusion MSAFF method; the MSAFF method includes a generator G stru and fusion device E fusion ; The generator G stru , used to transform the secret information tensor Z M Encoded into the preliminary structure tensor S related to the image structure information tensor M middle; The fusion device E fusion , used to combine the image structure information tensor S obtained in step S21 and the generator G stru The generated preliminary structure tensor S M Fusion is performed to obtain enhanced structural information tensor S23, introduce a stego image generator G based on a GAN network, and generate different images by inputting different combinations of structure information tensors and texture information tensors T; the combination of structure information tensors and texture information tensors T includes the original structure information tensor S and texture information tensor T, the preliminary structure tensor S M With texture information tensor T, enhanced structure information tensor With the texture information tensor T S24, introduce the discriminator D co and a discriminator D; the discriminator D co At the same time, the texture loss between the image generated in step S23 and the input image I is calculated; the discriminator D is used to optimize the steganographic image generator G to ensure the authenticity and naturalness of the generated image.

4. The non-embedding image zero-distortion steganography method according to claim 3, characterized in that: In step S21, during encoder training, the discriminator D is introduced dist Encourage the texture information tensor T to follow a specific distribution, then the loss function L during training E,dist is defined as: L E,dist =softplus(-D dist (T)); The softplus function is defined as: softplus(x)=log(1+e x ).

5. The non-embedding image zero-distortion steganography method according to claim 4, characterized in that: In step S22, the fusion device E fusion The specific process includes the following: S221 uses the multi-scale attention feature fusion MSAFF method to calculate the fused weight value to enhance the expression ability of structured information. The specific formula is expressed as: Among them, σ represents the sigmoid activation function; MLP represents the multi-layer perceptron; AvgPool represents the average pooling; MaxPool represents the maximum pooling; Indicates the initial feature integration. In order to simplify the calculation, the element-level summation is selected as the initial integration, and the calculation result is Denoted as symbol α; then the enhanced structural information for: in, represents element-wise product; Represents the final structured information of the stego-image; Represents the real space, C represents the number of channels, H represents the number of pixels in the height of the input image, and W represents the number of pixels in the width of the input image; the weight value after fusion That is, α is a real number between 0 and 1, which is used to control the network between S and S M Parameter selection ratio between ; S222, introduce a discriminator D dist_struc , so that the fused structured information The distribution of is approximately equal to S, then the corresponding loss function is defined as: final, The feature distribution of will tend to the feature distribution of the real image.

6. The non-embedding image zero-distortion steganography method according to claim 5, characterized in that: Step S23 also includes the following steps: S231, based on the combination of different input structural information tensors and texture information tensors T, three different stego images are obtained, namely: I1=G(S,T); Among them, I2 and I3 are steganographic images containing secret information, and the corresponding loss function L G,rec as follows: L G,rec =||I1-I|| 1 ; Among them, ||·|| 1 Represents L1 loss, which is the absolute loss function.

7. The non-embedding image zero-distortion steganography method according to claim 6, characterized in that: Step S24 includes the following steps: S241, using the discriminator D co Compute the texture loss between I2 and I; discriminator D co Cut out an image block from each of the stego image I2 and the input image I, encode and connect the cut image blocks, and use the network to calculate the loss L during training. G,texture : L G,texture =D co (patch(I2),patch(I)); Among them, the function Patch(·) represents the cropping function; S242, introduce the discriminator D to optimize the generator G and the overall loss function L G,real as follows: L G,real =D(I1,I)+D(I2,I)+D(I3,I)。 8. The non-embedding image zero-distortion steganography method according to claim 7, characterized in that: Step S3 includes the following steps: S31, the extractor E extracts the corresponding structural information from the stego-image I2 or I3 Right now: The loss function L when training the extractor E E,struc for: The extractor E shares parameters with the encoder E; S32, Extractor E fusion,ex Further from structured information Extract the tensor corresponding to the secret information M from Right now: Training the extractor E fusion,ex The loss function L when E,fusion for: S33, another extractor E x from Extracting secret information Right now: Training the extractor E x The loss function L when Ex for:

9. A non-embedding image zero-distortion steganography system, used to implement the non-embedding image zero-distortion steganography method according to any one of claims 1 to 8, characterized in that: The non-embedding image zero-distortion steganography system comprises: A secret information processing module, used for performing tensor processing on secret information; A stego image generation module, used to generate a stego image based on the secret information after tensorization processing; The secret information extraction module is used to re-extract the secret information from the generated stego image.

Images