Depth image steganography method and device based on black box generation model

Abstract

The invention discloses a depth image steganography method and device based on a black box generation model, and the method comprises the steps: firstly generating a carrier image and a fluctuation image through the black box generation model, screening an effective fluctuation image through a mean square error estimation function MSE threshold value, and guaranteeing that the similarity of the screened fluctuation image and the carrier image meets the requirement; constructing an encoder and a decoder, and constraining the encoder through a multi-loss function to generate a steganographic image; through confrontation training of the discriminator and the encoder, the generated steganographic image accords with fluctuation distribution of the fluctuation image generated by the black box generation model, and the anti-detection capability of the steganographic image is improved. According to the method, the steganographic image is disguised as model natural fluctuation by mining the parameter fluctuation rule from the text to the image generation model, so that the detection resistance and generalization ability of the steganographic image are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image steganography technology, and in particular to a deep image steganography method and apparatus based on a black-box generation model. Background Technology

[0002] Image steganography is a core technology in the field of covert communication. Its core objective is to embed secret information into a carrier image, making the steganographic image visually indistinguishable from the carrier image, while ensuring that the communication itself remains undetected. Unlike encryption technology, which only protects the content, steganography further conceals the fact of the "communication behavior." Based on their design methods, existing image steganography methods can be divided into three categories: manual design, semi-machine design, and fully machine design.

[0003] Throughout the development of steganography, manually designed methods have long dominated. These methods use natural images as a medium and achieve information hiding by slightly modifying pixels. For example, adaptive steganography algorithms (such as WOW and UNIWARD) select modification regions based on image texture complexity, while minimal distortion steganography (such as Syndrome-Trellis codes (STC)) improves security by optimizing the distortion function. However, the distribution of natural images is highly complex and difficult to model; even minor steganographic modifications can be detected by steganography analysis tools, thus limiting security.

[0004] In recent years, the development of AI technology has driven the evolution of steganography towards semi-machine design. Researchers have designed distortion functions based on Generative Adversarial Networks (GANs) (such as ASDLGAN and UT-GAN), using GANs to learn image modification probability maps to replace traditional manually designed distortion functions. However, these methods are essentially still minimal-distortion steganography, relying on embedding messages with steganographic codes, and have the inherent limitation of low message capacity.

[0005] While end-to-end steganography methods designed entirely by machines can improve message capacity, they have obvious drawbacks: "Hidden image" methods (such as HiNet and ISSN) embed secret images into the carrier, which is not suitable for binary messages; "Hidden binary message" methods (such as ChatGAN) support binary messages, but steganographic modifications are prone to leaving statistical traces, making it difficult to balance extraction accuracy and anti-detection capabilities in the face of deep learning steganalysis tools (such as XuNet and CovNet).

[0006] In summary, existing steganography methods suffer from three major drawbacks in black-box generative model scenarios: first, the lack of a "natural noise" modeling method adapted to black-box models; second, the tendency for steganographic images to deviate from the distribution of generated images; and third, insufficient resistance to steganalysis. Therefore, there is an urgent need for an image steganography method that is both secure and practical for black-box generative models. Summary of the Invention

[0007] The purpose of this invention is to provide a deep image steganography method and apparatus based on a black-box generation model. This method, by mining the parameter fluctuation patterns of the text-to-image generation model, disguises the steganographic image as a natural fluctuation of the model, effectively improving the anti-detection and generalization ability of the steganographic image.

[0008] The objective of this invention is achieved through the following technical solution: A depth image steganography method based on a black-box generative model, the method comprising: Step 1: Use a black-box generation model to generate carrier images and wave images, and use the mean square error estimation function (MSE) threshold to filter valid wave images to ensure that the similarity between the filtered wave images and carrier images meets the requirements. Step 2: Construct the encoder and decoder, and use multiple loss functions to constrain the encoder to generate steganalytic images; Step 3: Through adversarial training between the discriminator and the encoder, the generated steganalysis image is made to conform to the fluctuation distribution of the fluctuation image generated by the black-box generation model, thereby improving the anti-detection capability of the steganalysis image.

[0009] A depth image steganography device based on a black-box generative model, the device comprising: The wave image generation unit is used to generate carrier images and wave images using a black-box generation model, and to filter valid wave images by using the mean square error estimation function (MSE) threshold to ensure that the similarity between the filtered wave images and carrier images meets the requirements. The encoder-decoder building unit is used to build the encoder and decoder, and generates steganalog images by constraining the encoder through multiple loss functions. The adversarial training unit is used to train the discriminator and encoder adversarially so that the generated steganases conform to the fluctuation distribution of the fluctuation images generated by the black-box generation model, thereby improving the anti-detection ability of the steganases.

[0010] As can be seen from the technical solution provided by the present invention, the above method, by mining the parameter fluctuation patterns of the text-to-image generation model, disguises the stegana image as a natural fluctuation of the model, effectively improving the anti-detection and generalization ability of the stegana image, which is of great significance for solving the stegana security problem of black-box generation model. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1A schematic diagram of the process of a depth image steganography method based on a black-box generation model provided in an embodiment of the present invention; Figure 2 Different examples are given in this invention. A schematic diagram illustrating the change in accuracy of steganalysis under different values; Figure 3 This is a schematic diagram comparing the steganalysis accuracy of DIFS and the baseline method in the example given in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments, and do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0014] like Figure 1 The diagram shows a flowchart of a depth image steganography method based on a black-box generation model provided in an embodiment of the present invention. The method includes: Step 1: Use a black-box generation model to generate carrier images and wave images, and use the mean square error estimation function (MSE) threshold to filter valid wave images to ensure that the similarity between the filtered wave images and carrier images meets the requirements. In this step, the process of generating carrier images and wave images using a black-box generation model is as follows: Let's define a black-box generation model. The input is semantic control (i.e., text description, such as "a dog") and parameters (For example, the CFG scale, used to control the alignment between the image and the text description), the output is a carrier image. With wave image , , The number of wave images corresponding to a single carrier; in, ; , indicating a color image; , indicating the image size; , indicating the range of pixel values; Control by fixed semantics Fine-tuning the black-box generation model parameters (For example, adjusting the CFG scale from 7.4980 to 7.4981, 7.4982, etc.), generating A wave image can be generated, for example, 10 wave images can be generated.

[0015] The process of filtering valid fluctuation images using the mean square error estimation function (MSE) threshold is as follows: Calculate each wave image With carrier image The mean square error estimation function MSE is retained only if it satisfies The wave image; in, Set the MSE screening threshold; MSE screening ensures that the selected wave images are visually similar to the carrier images, thus avoiding the introduction of additional statistical bias.

[0016] Experiments show that the MSE screening threshold It can achieve the best balance between image quality and concealment.

[0017] In practice, by generating fluctuating images in batches, the output distribution of the model can be backfitted, providing guidance for the steganalysis and preventing the steganalysis from deviating from the statistical characteristics of the generation model.

[0018] Step 2: Construct the encoder and decoder, and use multiple loss functions to constrain the encoder to generate steganalytic images; In this step, the process of building the encoder and decoder is as follows: The encoder uses a U-Net architecture, with the input being a carrier image. With secret messages The splicing features, with dimensions of Multi-scale features are extracted through 4-layer downsampling and 4-layer upsampling to output a steganalytic image. Among them, secret messages are set. , Indicates load (unit: bpp). Indicates image size, steganography By encoder Generate, satisfying: ; For encoder The parameters, and the pixel modification magnitude between the steg image and the carrier image is subject to the truncation factor. Constraints, namely: ,have ; where subscript This represents the spatial dimension of an image pixel, specifically its height. ,width Position index on; In practice, a truncation operation is added to the output layer to ensure that the pixel modification range between the stegana and the carrier image does not exceed [the specified value]. ; decoder Also using the U-Net architecture, the input is a steganalyte image. Features are extracted through downsampling and upsampling layers symmetrical to the encoder, resulting in an output dimension of... Extraction message ,satisfy: ; For decoder Parameters; Add a sigmoid activation function to the output layer to map the output value to the [0,1] interval, and obtain the binary extraction result through rounding.

[0019] The multiple loss function is specifically defined as follows: Image distortion loss The steganalysis image is calculated using the mean square error estimation function (MSE). With carrier image The difference ensures visual indistinguishability; the formula is: ; Losses from message distortion : Calculate secret messages using binary cross-entropy With Retrieval Message To ensure accurate extraction, the formula is: ; in , Indicates secret message With Retrieval Message The corresponding index is , , Element; As the first dimension, ; As the second dimension, ; As the third dimension, ; Image fluctuation loss Randomly select one wave image from the effective wave image set F. The steganalysis image is calculated using the mean square error estimation function (MSE). With wave image The difference ensures that the stegtext image fits the fluctuation distribution, and the formula is: ; Total loss The loss terms are balanced using weighting coefficients, as shown in the formula: ; in , , These are the weighting coefficients corresponding to each loss term.

[0020] In practice, experiments have verified that... , , The optimal weighting coefficients are...

[0021] By aggregating multi-dimensional constraints to enhance steganography robustness and transferability, and learning the overall fluctuation patterns of generative models, steganography transfer can be achieved across model scenarios.

[0022] Step 3: Through adversarial training between the discriminator and the encoder, the generated steganalysis image is made to conform to the fluctuation distribution of the fluctuation image generated by the black-box generation model, thereby improving the anti-detection capability of the steganalysis image.

[0023] In this step, the discriminator aims to distinguish the generated images. With Stegogram That is, the carrier image Wave image Output low score (low probability) for stegographic images Output high score (high probability); Among them, the discriminator It employs a Convolutional Neural Network (CNN) architecture, with the input being a carrier image. Stegographic images Wave image The output is the classification score. ,satisfy: ; These are the parameters for the discriminator; For input items, ; The discriminator uses ReLU activation function in its convolutional layers and sigmoid activation function in its fully connected layers to ensure that the output value is in the interval [0,1]. Let the output probability of the discriminator be... Then the discriminator loss function Defined as: .

[0024] In practice, an alternating training strategy can be adopted: In each training round, the parameters of the encoder and decoder are fixed first, and the parameters of the discriminator are updated (e.g., iterate 5 times). Then, fix the parameters of the discriminator and update the parameters of the encoder and decoder (e.g., iterate once) to ensure dynamic balance between the two, so that the steganalysis image is statistically indistinguishable from the generated image.

[0025] The above scheme utilizes two key characteristics of black-box generative models: 1. Fine-tuning of input parameters can produce visually similar but pixel-distributed fluctuating images; 2. The generated results must conform to the model's own statistical characteristics. Based on this, a corresponding steganography camouflage strategy is designed. By constraining the steganalysis image to simultaneously match the distribution of both the carrier image and the fluctuating image, the steganographic modification is disguised as a natural fluctuation of the model, effectively preventing detection by steganalysis tools. This ensures that the steganalysis image not only conforms to the visual features of the carrier image but also matches the distribution characteristics of the fluctuating image, achieving both "visual indistinguishability" and "statistical undetectability."

[0026] Based on the above method, embodiments of the present invention also provide a depth image steganography device based on a black-box generation model, the device comprising: The wave image generation unit is used to generate carrier images and wave images using a black-box generation model, and to filter valid wave images by using the mean square error estimation function (MSE) threshold to ensure that the similarity between the filtered wave images and carrier images meets the requirements. The encoder-decoder building unit is used to build the encoder and decoder, and generates steganalog images by constraining the encoder through multiple loss functions. The adversarial training unit is used to train the discriminator and encoder adversarially so that the generated steganases conform to the fluctuation distribution of the fluctuation images generated by the black-box generation model, thereby improving the anti-detection ability of the steganases.

[0027] The specific implementation process of each unit in the above device is described in the method embodiment.

[0028] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method.

[0029] This invention also provides a computer storage medium storing a plurality of instructions adapted for loading and executing the method by a processor.

[0030] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0031] To illustrate the effectiveness of the method described in the embodiments of the present invention, specific experiments are conducted below: I. Experimental Setup To verify the effectiveness of the deep image steganography method based on a black-box generative model (hereinafter referred to as DIFS) described in this invention, the experiment uses the current mainstream text-to-image generation model, namely the black-box generative model Stable Diffusion, to construct the experimental dataset. The specific settings are as follows: (I) Dataset Construction Using 1000 categories from the ImageNet dataset as semantic control input, the Stable Diffusion model generates images. For each category, 10 different random seeds are used to generate images, resulting in 10,000 color images of size 512×512 pixels. The Stable Diffusion input parameter "CFG scale" (controlling the alignment between images and text descriptions) is fine-tuned to construct a coverage dataset and a fluctuation dataset: the CFG scale is sequentially fine-tuned from 7.4980 to 7.4981, 7.4982, etc., generating 11 datasets, each with 10,000 images. One of these serves as the carrier dataset, and the remaining 10 as fluctuation datasets. The images in the fluctuation datasets correspond one-to-one with those in the carrier datasets to ensure consistency in size and content. During the model training phase, the carrier dataset is randomly divided into a training set (9000 images) and a validation set (1000 images). For the DIFS model, each fluctuation dataset is also divided into a training subset (9000 images) and a validation subset (1000 images) in a 9:1 ratio to ensure data independence in the training and validation processes.

[0032] (II) Hardware and Software Environment The experimental hardware platform uses a computer equipped with an NVIDIA GEFORCE RTX 2080 Ti graphics card, and the software environment is built on the PyTorch deep learning framework. All model training and testing are completed in this environment to ensure the stability and reproducibility of the experimental process.

[0033] II. Evaluation Indicators To comprehensively evaluate the performance of DIFS in this application, evaluation metrics were set from three dimensions: image quality, accuracy of secret information extraction, and resistance to steganalysis, as follows: 1. Image quality evaluation indicators Mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) are used to quantify the differences between stegographic images and overlaid or fluctuating images, reflecting the visual quality of the images. Specifically: Mean Squared Error (MSE): Used to measure the average squared difference between pixel values ​​of two images. The calculation formula is as follows: ; Where N is the total number of pixels in the image; X(i,j) and Y(i,j) are the pixel values ​​at position (i,j) in the two images, respectively. The smaller the MSE value, the smaller the difference between the two images, and the better the image quality.

[0034] Peak Signal-to-Noise Ratio (PSNR): Calculated based on MSE, used to evaluate the degree of image distortion, measured in decibels (dB). The calculation formula is as follows: ; Where MAX is the maximum possible value of a pixel (MAX=255 for an 8-bit grayscale image). The higher the PSNR value, the less image distortion and the higher the quality.

[0035] Structural Similarity Index (SSIM): Measures image similarity from three dimensions: brightness, contrast, and structure. The value ranges from [0,1]. The calculation formula is as follows: ; in, , These are the pixel mean values ​​for images X and Y, respectively; , , respectively, are the pixel variances of images X and Y; Let X be the covariance of the images X and Y. , SSIM is a stability constant. The closer the SSIM value is to 1, the more similar the structure and visual effect of the two images are, and the better the quality.

[0036] 2. Accuracy of extracting secret information Since the decoder output is a floating-point number in the range [0,1], it is converted to a binary value (0 or 1) by rounding. The bit consistency rate between the extracted secret information and the original secret information is calculated, which is the extraction accuracy. The higher the accuracy, the stronger the reliability of the secret information transmission.

[0037] 3. Resistance to steganalysis The classification accuracy of the steganalysis network on the test set is used as the evaluation metric, and the calculation formula is as follows: ; in, The number of samples correctly predicted by the steganography analysis network; The total number of samples in the test set is given. A lower accuracy indicates that the steganalyte is more difficult to detect, and the stronger DIFS's resistance to steganalysis. The experiments used the currently mainstream steganalysis networks CovNet and LWENet as test models to ensure the objectivity and representativeness of the evaluation results.

[0038] III. Experiment and Results Analysis 1. Hyperparameter optimization experiment Volatility Loss Weight :when At that time, DIFS achieved an extraction accuracy of 99.45%, PSNR=46.205 dB, and SSIM=0.9939; while CovNet's detection accuracy was as low as 82.29%, which was superior to [the previous method]. (Detection rate 86.25%) and (Extraction rate 96.95%), therefore selected .

[0039] Table 1 Image Fluctuation Loss Weights Performance comparison

[0040] Wave image filtering threshold :like Figure 2 The examples shown are different from those given in this invention. A diagram illustrating the change in steganalysis accuracy under different values, when... At that time, DIFS had the lowest steganalysis accuracy (CovNet=82.29%); therefore, it was chosen. .

[0041] 2. Anti-steganography performance test Two steganalysis tools, CovNet and LWENet, were tested at different training set sizes (25-300 pairs), such as... Figure 3 The diagram shown is a comparison of the steganalysis accuracy of the example DIFS and the baseline method described in this invention (using...). For example): When the training set has 300 pairs, the detection rate of CovNet on (c,s) pairs in this application's DIFS is 82.29%, which is 14% lower than that of ChatGAN (96.29%) in the prior art; the detection rate of LWENet is 75.37%, which is 7.34% lower than that of ChatGAN (82.71%). Furthermore, the CovNet detection rate of DIFS in this application was further reduced to 80.89%, proving the effectiveness of the fluctuation camouflage strategy.

[0042] 3. Safety Analysis Experiment The security of DIFS in this application was tested for two threat models: "test data known" (scenario 0) and "test data unknown" (scenarios 1-3). The results are shown in Table 2 below: Table 2. Steganography accuracy (%) under hyperparameter confidentiality conditions

[0043] Scenario 1 (Attacker unknown) ): The CovNet detection rate of DIFS decreased to 70.97%, which is 25.32% lower than that of ChatGAN (96.29%); Scenario 2 (Attacker has no known random seed): DIFS's CovNet detection rate drops to 76.60%, a decrease of 14.30% compared to ChatGAN (90.90%); Scenario 3 (Attacker unknown) The detection rate of DIFS further decreased to 70.62%, proving that hyperparameter confidentiality can significantly improve security.

[0044] 4. Comparison with other steganography methods Compared with image steganography methods (such as HiNet, ISSN, and CRoSS), HiNet and ISSN have PSNR < 30dB, indicating significant noise in the steganographic images; CRoSS has a message extraction accuracy of only 53.7%, and the generated content deviates from the text prompt; the PSNR of DIFS in this application is > 46dB, and the extraction accuracy is 99.45%, showing significantly better performance in binary message steganography scenarios.

[0045] 5. The effect of load on performance The performance of DIFS under different loads (1-3 bpp) was tested, and the results are shown in Table 3: Table 3 Performance indicators of DIFS under different loads

[0046] When the load increased from 1 bpp to 3 bpp, the PSNR decreased from 46.205 dB to 43.883 dB, and the extraction accuracy decreased from 99.45% to 96.89%, but it still remained at a practical level. The CovNet detection rate increased from 82.29% to 97.21%, so in practical applications, the payload should be selected according to security requirements.

[0047] 6. Generalization test DIFS and ChatGAN were trained on the IN dataset and tested on DDB (user-generated images) and MD (Midjourney images). The results are shown in Table 4. Table 4 Generalization tests (%) on different datasets

[0048] ChatGAN achieved CovNet detection rates of 98.93% and 99.01% on DDB and MD respectively, but exhibited poor generalization. The detection rate of DIFS in this application decreased to 76.94% on DDB and remained at 86.50% on MD, demonstrating that it has the ability to generalize across datasets by learning the fluctuating distribution of the generative model.

[0049] In summary, the Deep Image Fluctuation Steganography (DIFS) method based on a black-box generative model described in this embodiment of the invention can convert the carrier image output by the black-box generative model into a steganalyte image. When a malicious user uses a steganalysis tool to detect the steganalyte image, the steganalyte image can evade the detection tool's identification because it conforms to the natural fluctuation distribution of the black-box generative model, while ensuring that legitimate users can accurately extract secret messages through the decoder.

[0050] The method described in this invention is adapted to black-box generation model scenarios and does not require obtaining the model's internal parameters; the stegated images have both visual indistinguishability and statistical concealment, and their resistance to steganalysis is significantly better than traditional methods; at the same time, the framework of this invention has good generalization ability and can maintain stable performance on different generated image datasets.

[0051] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0052] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A depth image steganography method based on a black-box generative model, characterized in that, The method includes: Step 1: Use a black-box generation model to generate carrier images and wave images, and use the mean square error estimation function (MSE) threshold to filter valid wave images to ensure that the similarity between the filtered wave images and carrier images meets the requirements. Step 2: Construct the encoder and decoder, and use multiple loss functions to constrain the encoder to generate steganalytic images; Step 3: Through adversarial training between the discriminator and the encoder, the generated steganalysis image is made to conform to the fluctuation distribution of the fluctuation image generated by the black-box generation model, thereby improving the anti-detection capability of the steganalysis image.

2. The image steganography security enhancement method based on the black-box generation model according to claim 1, characterized in that, In step 1, the process of generating the carrier image and wave image using the black-box generation model is as follows: Let's define a black-box generation model. The input is semantic control With parameters The output is a carrier image. With wave image , , The number of wave images corresponding to a single carrier; in, ; , indicating a color image; , indicating the image size; , indicating the range of pixel values; Control by fixed semantics Fine-tuning the black-box generation model parameters ,generate A wave pattern.

3. The image steganography security enhancement method based on the black-box generation model according to claim 2, characterized in that, In step 1, the process of filtering valid fluctuation images using the mean square error estimation function (MSE) threshold is as follows: Calculate each wave image With carrier image The mean square error estimation function MSE is retained only if it satisfies The wave image; in, Set the MSE screening threshold; MSE screening ensures that the selected wave images are visually similar to the carrier images, thus avoiding the introduction of additional statistical bias.

4. The image steganography security enhancement method based on the black-box generation model according to claim 1, characterized in that, In step 2, the process of constructing the encoder and decoder is as follows: The encoder uses a U-shaped network architecture, with the input being a carrier image. With secret messages The splicing features, with dimensions of Multi-scale features are extracted through 4-layer downsampling and 4-layer upsampling to output a steganalytic image. Among them, secret messages are set. , Indicates load, Indicates image size, steganography By encoder Generate, satisfying: ； For encoder The parameters, and the pixel modification magnitude between the steg image and the carrier image is subject to the truncation factor. Constraints, namely: ,have ; where subscript This represents the spatial dimension of an image pixel, specifically its height. ,width Position index on; decoder Also using the U-Net architecture, the input is a steganalyte image. Features are extracted through downsampling and upsampling layers symmetrical to the encoder, resulting in an output dimension of... Extraction message ,satisfy: ； For decoder Parameters; Add a sigmoid activation function to the output layer to map the output value to the [0,1] interval, and obtain the binary extraction result through rounding.

5. The image steganography security enhancement method based on the black-box generation model according to claim 4, characterized in that, In step 2, the multiple loss function is specifically defined as follows: Image distortion loss The steganalysis image is calculated using the mean square error estimation function (MSE). With carrier image The difference ensures visual indistinguishability; the formula is: ; Losses from message distortion : Calculate secret messages using binary cross-entropy With Retrieval Message To ensure accurate extraction, the formula is: ； in , Indicates secret message With Retrieval Message The corresponding index is , , Element; As the first dimension, ; As the second dimension, ; As the third dimension, ; Image fluctuation loss Randomly select one wave image from the effective wave image set F. The steganalysis image is calculated using the mean square error estimation function (MSE). With wave image The difference ensures that the stegtext image fits the fluctuation distribution, and the formula is: ; Total loss The loss terms are balanced using weighting coefficients, as shown in the formula: ； in , , These are the weighting coefficients corresponding to each loss term.

6. The image steganography security enhancement method based on the black-box generation model according to claim 5, characterized in that, In step 3, the discriminator aims to distinguish the generated images. With Stegogram That is, the carrier image Wave image Low output score for stegographic images Output high scores; Among them, the discriminator A convolutional neural network architecture is used, with the input being a carrier image. Stegographic images Wave image The output is the classification score. ,satisfy: ； These are the parameters for the discriminator; For input items, ; The discriminator uses ReLU activation function in its convolutional layers and sigmoid activation function in its fully connected layers to ensure that the output value is in the interval [0,1]. Let the output probability of the discriminator be... Then the discriminator loss function Defined as: 。 7. The image steganography security enhancement method based on the black-box generation model according to claim 6, characterized in that, In step 3, an alternating training strategy is specifically adopted: In each training round, the parameters of the encoder and decoder are fixed first, and the parameters of the discriminator are updated. Then, fix the parameters of the discriminator, update the parameters of the encoder and decoder, ensure dynamic balance between the two, and finally make the steganalysis image statistically indistinguishable from the generated image.

8. A depth image steganography device based on a black-box generative model, characterized in that, The device includes: The wave image generation unit is used to generate carrier images and wave images using a black-box generation model, and to filter valid wave images by using the mean square error estimation function (MSE) threshold to ensure that the similarity between the filtered wave images and carrier images meets the requirements. The encoder-decoder building unit is used to build the encoder and decoder, and generates steganalog images by constraining the encoder through multiple loss functions. The adversarial training unit is used to train the discriminator and encoder adversarially so that the generated steganases conform to the fluctuation distribution of the fluctuation images generated by the black-box generation model, thereby improving the anti-detection ability of the steganases.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the method of any one of claims 1 to 7.

Images