A large-capacity steganographic method for color images based on a carrier generation method based on a diffusion model

Abstract

The present invention discloses a large-capacity steganographic method for color images based on a carrier generation type based on a diffusion model, which relates to the field of image processing technology and is characterized in that it includes the following steps: S1: training a diffusion model separately, allowing the diffusion model to fully learn the image semantics of a given data set and sample images with the same distribution as the given data set, and use them as carrier images of the image steganographic part; S2: in the image steganographic part based on FusionUnet. The technical problem to be solved by the present invention is to provide a large-capacity steganographic method for color images based on a carrier generation type based on a diffusion model, embedding a secret image into a new carrier image generated by sampling a diffusion model, so that the steganographic image obtained by embedding on the generated new carrier image not only has extremely high visual quality, but also has stronger anti-steganographic analysis capability.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a carrier generation type color image large-capacity steganographic method based on a diffusion model. Background Art

[0002] Steganography refers to embedding confidential information into the inconspicuous properties of the carrier medium, which is an effective method for transmitting secret information. Since images contain a large amount of redundant information, images become ideal carriers for steganography. The main task of image steganography is to embed more information into the image while reducing image distortion. Among them, the STC (syndrome trelliscode) algorithm is one of the most efficient steganographic encoding methods. Under a given payload, the image distortion it produces is minimal, even close to the theoretical lower limit. Therefore, after the emergence of the STC steganographic encoding algorithm, the core work of image steganography is to design a reasonable distortion function.

[0003] Image steganography algorithms can be roughly divided into heuristic-based steganography and deep learning-based steganography. For heuristic-based steganography algorithms, such as LSB, HUGO, WOW, S-UNIWARD, Hill, and MiPOD, the pixel values ​​of the carrier image are generally modified according to a pre-designed constraint function to embed secret information. Specifically, the S-UNIWARD and WOW algorithms minimize the embedding distortion and improve the visual quality of the stego image by selecting appropriate modification positions to avoid embedding in smooth areas.

[0004] In recent years, with the development of deep learning technology, convolutional neural network (CNN) has been widely used in the field of steganography to improve steganographic performance due to its powerful feature learning ability. For example, Hayes et al. designed an adaptive steganography algorithm using adversarial thinking, making two networks compete with each other to generate steganographic images. Baluja et al. constructed multiple deep neural networks for image preprocessing and data hiding, respectively, thereby effectively improving the quality of steganographic images. At the same time, steganographic models based on generative adversarial networks (GAN) have also received increasing attention. Tang et al. implemented adaptive calculation of embedding distortion to avoid embedding information into smooth areas of the image, and introduced XuNet as a steganalyzer for adversarial training, which enhanced the imperceptibility of the steganographic image. Yang et al. further proposed a simple distortion calculation function to form a new steganographic framework. By replacing the TES neural network proposed by Tang with a double tanh activation function that does not require pre-training, its adversarial training time is significantly reduced compared to the steganographic algorithm based on ASDL-GAN.

[0005] As a generative model that has attracted much attention in recent years, the diffusion model has a more powerful image generation capability than the generative adversarial network (GAN). The execution process of the diffusion model mainly includes two parts: the forward denoising process and the reverse denoising process. In the forward denoising process, noise is gradually added to the original image to obtain a noisy image that obeys the Gaussian distribution. The reverse denoising process gradually removes the noise and finally obtains an image that obeys the distribution of the original image. The random sampling algorithm (DDPM) algorithm uses the diffusion model to generate images with high resolution and is widely used. Its disadvantage is that the image generation steps are many and time-consuming. Therefore, in order to improve the sampling efficiency, the deterministic sampling algorithm (DDIM) proposed by Song et al. shortens the sampling steps by using non-Markov chains, effectively shortening the image sampling time.

[0006] The prior art has the following deficiencies:

[0007] 1. Common deep learning-based image steganography algorithms all select carrier images based on existing data sets to embed secret images. The original carrier images will be distorted after a large amount of secret information is embedded.

[0008] 2. The degree of feature fusion between the carrier image and the secret image has a direct impact on image steganography. The existing technology has the problem of feature map information loss;

[0009] 3. Anti-steganalysis capability, as an important indicator for judging steganographic algorithms, also directly reflects the security of steganographic images. The existing technology has poor anti-detection capability;

[0010] 4. For image steganography schemes based on deep learning, the quality of loss function design is directly related to the efficiency and effect of model training. The effect of the steganography part of the existing technology is not good. Summary of the invention

[0011] The technical problem to be solved by the present invention is to provide a carrier-generated color image large-capacity steganography method based on a diffusion model, in which a secret image is embedded into a new carrier image generated by sampling the diffusion model. Thanks to the dual feedback of the steganography part and the steganalysis part to the carrier generation part, the steganalysis image obtained by embedding on the generated new carrier image not only has extremely high visual quality, but also has stronger anti-steganalysis ability, thus achieving excellent performance in large-capacity steganography.

[0012] The present invention adopts the following technical solutions to achieve the invention objectives:

[0013] The carrier generation type color image large-capacity steganography method based on the diffusion model is characterized by comprising the following steps:

[0014] S1: Train the diffusion model separately, so that the diffusion model can fully learn the image semantics of the given dataset and sample images with the same distribution as the given dataset, which will be used as the carrier images of the steganographic part of the image;

[0015] S2: In the image steganography part based on FusionUnet, the sampled carrier image and the secret image are stacked according to the channel dimension and input into the generator built based on FusionUnet. After the convolution operation and feature fusion of the convolution layer, the steganographic image is output. The steganographic image is visually consistent with the generated carrier image.

[0016] S3: The generated steganalysis image is input into the steganalysis adversarial network for discrimination by the steganalyzer, and the discrimination loss is generated and fed back to the generator and diffusion model in the image steganalysis part together with the mean square error and structural similarity for optimization. The steganalysis effect is gradually improved through continuous iteration of the model.

[0017] S4: The FusionUnet-based extractor realizes the extraction of the embedded secret image, that is, obtaining an extracted image that is visually consistent with the secret image, thereby completing the image steganography and extraction process.

[0018] As a further limitation of the technical solution, constructing a diffusion model includes a carrier image generation part based on the diffusion model, an image steganography part, an image extraction part and a steganalysis confrontation part;

[0019] The image steganography part mainly includes a generator based on FusionUnet, which mainly consists of two parts: encoding part and decoding part;

[0020] The image extraction part mainly includes a FusionUnet-based extractor, which has the same structure as the generator;

[0021] The steganalysis adversarial part includes the steganalysis adversarial network of the commonly used spatial domain steganalyzer SRNet.

[0022] As a further limitation of the present technical solution, the encoder part consists of a dry convolution block and four downsampling blocks, each downsampling block contains two convolution layers and a maximum pooling layer, and the size of the feature map changes to half of the previous layer each time it passes through a downsampling block, and the number of channels doubles compared to the previous layer. The outputs of the dry convolution block and the first three downsampling blocks are input into the fusion module to perform feature fusion operations;

[0023] The decoder contains four upsampling blocks, which perform upsampling through bilinear interpolation operations and connect the feature maps obtained by the fusion module to finally restore the feature maps to the size of the original image.

[0024] As a further limitation of the present technical solution, the loss function includes a loss function of the image steganography part and a loss function of the carrier image generation part.

[0025] As a further limitation of the present technical solution, the specific steps of the loss function of the image steganography part are:

[0026] S51: Use mean square error loss to measure the difference between the stego image and the carrier image:

[0027] (1)

[0028] in: is the steganalysis loss between the stego-image and the carrier image;

[0029] is the pixel value of the carrier image;

[0030] is the pixel value of the stego-image;

[0031] is the number of pixels of the image;

[0032] S52: In order to narrow the gap between the extracted image and the original secret image, the model is trained using a loss that optimizes the visual similarity between the extracted image and the original embedded secret image. , which is calculated as follows:

[0033] (2)

[0034] in: is the pixel value of the carrier image;

[0035] is the pixel value of the stego-image;

[0036] is the number of pixels of the image;

[0037] S53: Based on the structural similarity loss, a loss is designed between the carrier image and the stego image. L STE2， The loss between the secret image and the extracted image L EXT2 ;

[0038] (3)

[0039] (4)

[0040] in: and is the mean and variance of the carrier image;

[0041] and is the mean and variance of the encrypted image;

[0042] and represents the mean and variance of the secret image;

[0043] and Indicates the mean and variance of the extracted image;

[0044] represents the covariance between the carrier image and the encrypted image;

[0045] represents the covariance between the secret image and the extracted image;

[0046] In the formula , L is the dynamic range of the image pixel values, K i is a coefficient much smaller than 1;

[0047] Based on the structural similarity loss function, it not only improves the similarity between the secret image and the carrier image, but also ensures the visual consistency between the extracted image and the secret image.

[0048] S54: In the steganalysis adversarial network, the steganalyzer discriminates the steganalyzer image and generates a discriminant loss L SAN， Improve the steganalysis resistance of the generated steganalysis image by feeding it back to the generator;

[0049] (5)

[0050] in: , is the Softmax output of the steganalyzer XuNet, , is the corresponding truth value;

[0051] S55: Total loss function of the generator of the image steganography part L S for:

[0052] (6)

[0053] The total loss function of the extractor in the image extraction part L E for:

[0054] (7).

[0055] As a further limitation of the technical solution, the steps of generating a partial loss function of the carrier image are:

[0056] S61: Using mean square error loss to measure the difference between images can effectively reduce the distance between two images. The calculation formula is defined as:

[0057] (8)

[0058] in: is the pixel value of the target image;

[0059] is the pixel value of the predicted image;

[0060] is the number of pixels of the image;

[0061] S62: Total loss function of carrier image generation based on diffusion model Total_DIF as follows:

[0062] (9)

[0063] in: It is the weight for adjusting the effect of loss.

[0064] Compared with the prior art, the advantages and positive effects of the present invention are:

[0065] 1. In the present invention, the diffusion model is first made to learn the data distribution of the original data set, and an image with the same distribution as the data set and more suitable for embedding is sampled from the noise through the deterministic sampling algorithm (DDIM) as the carrier image during steganography, thereby effectively solving the image distortion problem caused by the embedding of a large amount of information and increasing the embedding capacity to 24bpp. The present invention proposes a steganographic framework based on FusionUnet, which avoids the loss of feature map information by sampling the feature map in the horizontal and vertical directions instead of the pooling operation used when changing the size of the feature map. Therefore, the feature maps of the carrier image and the secret image are effectively fused to generate a high-quality steganographic image. The present invention effectively improves the anti-detection ability of the generated steganographic image by designing a steganalysis adversarial network for the carrier generation part based on the diffusion model and the image steganographic part. The present invention makes the distribution of the image generated by the diffusion model sampling approximate to the distribution of the original data set through the mean square error loss (MSE), so that it can sample images with the same semantics. In the steganographic part, the embedding process of the secret image is optimized by combining MSE, SSIM and steganalysis adversarial loss as the total loss, so that the steganographic image and the carrier image are as visually similar as possible. This loss is synchronously fed back to the diffusion model to guide it to generate a carrier image that is more suitable for embedding when sampling noise, thereby improving the effect of the steganographic part.

[0066] 2. This paper proposes a carrier generation method for large-capacity color image steganography based on semantic pre-learning diffusion model. The proposed scheme first learns the distribution of a given data set through a diffusion model and samples images with similar semantics as generated carrier images. Secondly, the secret image is embedded into the generated carrier image through the image steganography framework built by FusionUnet to obtain a steganographic image. The model generation process is optimized by the loss function designed based on the combined mean square error (MSE), structural similarity (SSIM) and adversarial loss. The dual feedback of the steganographic part and the steganalysis adversarial part to the diffusion model enables it to generate a carrier image that is more suitable for embedding. Experimental results show that compared with other advanced methods, this method embeds the secret image into the new carrier image sampled and generated by the diffusion model. Thanks to the dual feedback of the steganographic part and the steganalysis part to the carrier generation part, the steganographic image obtained by embedding on the generated new carrier image not only has extremely high visual quality, but also has stronger anti-steganography analysis ability, so it has excellent performance in large-capacity steganography. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0068] A specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation.

[0069] The present invention comprises the following steps:

[0070] S1: Train the diffusion model separately, so that the diffusion model can fully learn the image semantics of the given dataset and sample images with the same distribution as the given dataset, which will be used as the carrier images of the steganographic part of the image;

[0071] S2: In the image steganography part based on FusionUnet, the sampled carrier image and the secret image are stacked according to the channel dimension and input into the generator built based on FusionUnet. After the convolution operation and feature fusion of the convolution layer, the steganographic image is output. The steganographic image is visually consistent with the generated carrier image.

[0072] S3: The generated stego image is input into the steganalysis adversarial network for discrimination by the steganalyzer, and the discrimination loss is generated and fed back to the generator and diffusion model in the image steganalysis part together with the mean square error and structural similarity (SSIM) for optimization. The stego effect is gradually improved through continuous iteration of the model.

[0073] S4: The FusionUnet-based extractor realizes the extraction of the embedded secret image, that is, obtaining an extracted image that is visually consistent with the secret image, thereby completing the image steganography and extraction process.

[0074] The construction of the diffusion model includes a carrier image generation part based on the diffusion model, an image steganography part, an image extraction part and a steganalysis confrontation part;

[0075] In the carrier image generation part, first, the diffusion model is semantically pre-trained through a given data set, and the diffusion model is gradually learned through the mean square error to sample images with the same semantic distribution as the data set. Secondly, the generation process of the diffusion model is guided by the dual feedback of the steganalysis part and the steganalysis part, so as to ensure that the generated image obeys the distribution of the given data set and make the sampled image more suitable for embedding the steganalysis part of the image, thereby generating a high-quality steganalysis image.

[0076] The image steganography part mainly includes a generator based on FusionUnet, which mainly consists of two parts: the encoder part and the decoder part;

[0077] The image extraction part mainly includes a FusionUnet-based extractor, which has the same structure as the generator; the extractor can achieve high-quality extraction of the embedded secret image.

[0078] The steganalysis adversarial part includes the steganalysis adversarial network of the commonly used spatial domain steganalyzer SRNet. The generated image is discriminated through this network and the discriminant loss is generated. The discriminant loss is fed back to the generator and the diffusion model respectively to gradually optimize the generation ability of the model. This improves the anti-steganalysis ability of the generated image.

[0079] The encoder part consists of a dry convolution block and four downsampling blocks. Each downsampling block contains two convolution layers and a maximum pooling layer. Each time the downsampling block passes through, the size of the feature map changes to half of the previous layer, and the number of channels doubles compared to the previous layer. The outputs of the dry convolution block and the first three downsampling blocks are input into the fusion module to perform feature fusion operations;

[0080] The decoder contains four upsampling blocks, which perform upsampling through bilinear interpolation operations and connect the feature maps obtained by the fusion module to finally restore the feature maps to the size of the original image.

[0081] The loss function includes the loss function of the image steganography part and the loss function of the carrier image generation part.

[0082] First, we use a sampling loss based on the mean squared error ( L DIF) Approximate the semantic distribution of the image in the dataset by the image sampled by the diffusion model, so that the diffusion model can gradually denoise and sample the carrier image that conforms to the semantic information of the given dataset from the random noise image;

[0083] Secondly, in the image steganography part, the steganalysis loss between the steganalysis image and the carrier image is designed ( L STE1 ) and the stego-image structural similarity loss ( L STE2 ), so that the FusionUnet-based generator generates stego images with high visual quality; similarly, in order to improve the quality of the extracted image, the extraction loss is set accordingly L EXT1 Similarity loss with extracted image structure ( L EXT2 );

[0084] Finally, the steganalysis image is steganalyzed through the steganalysis adversarial network to generate the discriminative loss ( L SAN ).

[0085] The specific steps of the loss function of the image steganography part are:

[0086] S51: Mean Squared Error (MSE) is a commonly used loss function, especially in regression problems, which measures the difference between the model's predicted value and the true value. The calculation formula of the mean squared error is the average of the squares of the difference between the predicted value and the true value. Image steganography is a technology that uses images to hide images. The purpose is to make the steganographic image after embedding the secret image visually similar to the carrier image. The mean squared error loss is used to measure the difference between the steganographic image and the carrier image:

[0087] (1)

[0088] in: is the steganalysis loss between the stego-image and the carrier image;

[0089] is the pixel value of the carrier image;

[0090] is the pixel value of the stego-image;

[0091] is the number of pixels of the image;

[0092] S52: In order to narrow the gap between the extracted image and the original secret image, the model is trained using a loss that optimizes the visual similarity between the extracted image and the original embedded secret image. , which is calculated as follows:

[0093] (2)

[0094] in: is the pixel value of the carrier image;

[0095] is the pixel value of the stego-image;

[0096] is the number of pixels of the image;

[0097] S53: Using only a single pixel-level mean square error (MSE) loss function often results in the generated image having similar pixel values, but the visual effect is not natural. The structural similarity (SSIM) loss can better capture important features such as edges and textures in the image by considering the structural information of the image, making the generated image look more natural. The present invention also designs a loss function based on the structural similarity loss between the carrier image and the stego image. L STE2 ; The loss between the secret image and the extracted image L EXT2 ;

[0098] (3)

[0099] (4)

[0100] in: and is the mean and variance of the carrier image;

[0101] and is the mean and variance of the encrypted image;

[0102] and represents the mean and variance of the secret image;

[0103] and Indicates the mean and variance of the extracted image;

[0104] represents the covariance between the carrier image and the encrypted image;

[0105] represents the covariance between the secret image and the extracted image;

[0106] In the formula , L is the dynamic range of the image pixel values, K i is a coefficient much smaller than 1;

[0107] Based on the structural similarity loss function, it not only improves the similarity between the secret image and the carrier image, but also ensures the visual consistency between the extracted image and the secret image.

[0108] S54: In the steganalysis adversarial network, the steganalyzer discriminates the steganalyzer image and generates a discriminant loss L SAN , by feeding it back to the generator to improve the anti-steganalysis ability of the generated steganalysis image;

[0109] (5)

[0110] in: , is the Softmax output of the steganalyzer XuNet, , is the corresponding truth value;

[0111] S55: Total loss function of the generator of the image steganography part L S for:

[0112] (6)

[0113] The total loss function of the extractor in the image extraction part L E for:

[0114] (7).

[0115] The steps of the carrier image generation part of the loss function are:

[0116] S61: During the training process of the diffusion model, the diffusion model first gradually adds noise to the input target image, generates a full-noise image, and then performs inverse denoising to generate a predicted image. Using the mean square error loss to measure the difference between images can effectively reduce the distance between the two images. The calculation formula is defined as:

[0117] (8);

[0118] in: is the pixel value of the target image;

[0119] is the pixel value of the predicted image;

[0120] is the number of pixels of the image;

[0121] S62: The steganographic loss generated by the steganographic part is combined with the loss of the image predicted by the diffusion model LDIF The weighted joint is used as the total loss of the diffusion model to guide the image generation process, so as to ensure that the diffusion model not only generates images with the same distribution as the given dataset, but also makes the generated images more suitable for image steganography, thereby enhancing the effect of the image steganography part. Total loss function of the carrier image generation part based on the diffusion model Total_DIF as follows:

[0122] (9);

[0123] in: In order to adjust the weight of the loss effect, the experiment sets The value of is 100. The purpose is to make the carrier image generated by the diffusion model more suitable for image steganography without deviating from the distribution of the given dataset when retraining by loading the pre-trained diffusion model weights.

[0124] The training process is as follows:

[0125] In this experiment, the cat dataset is divided into two subsets, trainA and trainB. Each subset contains 15,000 color kitten images of size 64×64, and the test set consists of 10,000 color kitten images of the same size. The learning rate of the Adam optimizer is fixed at 1×10 -5 , batchsize is set to 30.

[0126] In order to avoid the problem of low training efficiency caused by large models and simultaneous training from scratch, the invention adopts a distributed training method in the training process. Figure 1 As shown, the data set is initially input into the diffusion model, and the diffusion model is trained by the loss LDIF, in order to enable the diffusion model to sample images with the same semantics as the data set. Subsequently, the image steganography, image extraction, and steganalysis adversarial parts are trained separately to obtain the weights of the pre-trained generator and extractor. After the two parts are trained separately, they are jointly refined. That is, the diffusion model first performs sampling, and inputs the sampled image into the image steganography part based on FusionUnet as a new carrier image, and then the steganography part embeds the secret image on this sampled new carrier image to realize the image steganography process. The generated steganalysis image is used as the input of the steganalysis adversarial network, and the steganalysis analyzer is used to discriminate the steganalysis image to generate a discriminant loss. L SAN , and in conjunction L S and L EThe model is optimized to ensure that the image generated by the diffusion model is more suitable for embedding the image steganography part, thereby enhancing the effect of image steganography.

[0127] This paper proposes a large-capacity steganographic method and system for color images based on a diffusion model. The carrier image generation part based on the diffusion model designed in the proposed scheme samples a carrier image that is more suitable for large-capacity information embedding, and the generated carrier image is input into the image steganographic part based on FusionUnet. The generator and extractor designed based on the FusionUnet structure complete the embedding and extraction of the secret image, thereby generating a steganographic image that is visually similar to the carrier image. The excellent image generation ability of the diffusion model ensures that the sampled image obeys the distribution of the given data set, and through the guidance of the loss of the joint steganographic part, it samples a carrier image that is more suitable for image steganography, thereby improving the steganographic ability of the carrier image. At the same time, the design of the steganalysis adversarial network discriminates the steganographic image to generate a discriminative loss, and the discriminative loss is fed back to the generation part to further improve the anti-steganographic analysis ability of the generated steganographic image. By designing a multiple loss function that combines mean square error loss, structural similarity loss and discriminative loss, it not only ensures the sampling effect of the diffusion model, but also ensures that the generated steganographic image has a very high visual similarity with the carrier image. Experimental results show that compared with other advanced methods, the proposed algorithm can obtain stegographic images with better visual quality when performing large-capacity steganography.

[0128] Table 1 Comparison of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of steganographic images and extracted images with other advanced schemes

[0129]

[0130] Table 2 Detection results of other advanced schemes under different commonly used steganalyzers

[0131]

[0132] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A carrier generation method for large-capacity color image steganography based on a diffusion model, characterized in that: The following steps are involved: S1: Train the diffusion model separately, so that the diffusion model can fully learn the image semantics of the given dataset and sample images with the same distribution as the given dataset, which will be used as the carrier images of the steganographic part of the image; S2: In the image steganography part based on FusionUnet, the sampled carrier image and the secret image are stacked according to the channel dimension and input into the generator built based on FusionUnet. After the convolution operation and feature fusion of the convolution layer, the steganographic image is output. The steganographic image is visually consistent with the generated carrier image. S3: The generated steganalysis image is input into the steganalysis adversarial network for discrimination by the steganalyzer, and the discrimination loss is generated and fed back to the generator and diffusion model in the image steganalysis part together with the mean square error and structural similarity for optimization. The steganalysis effect is gradually improved through continuous iteration of the model. S4: The FusionUnet-based extractor extracts the embedded secret image, that is, obtains an extracted image that is visually consistent with the secret image, thereby completing the image steganography and extraction process; The construction of the diffusion model includes a carrier image generation part based on the diffusion model, an image steganography part, an image extraction part and a steganalysis confrontation part; The image steganography part mainly includes a generator based on FusionUnet, which mainly consists of two parts: encoding part and decoding part; The image extraction part mainly includes a FusionUnet-based extractor, which has the same structure as the generator; The steganalysis adversarial part includes the steganalysis adversarial network of the commonly used spatial domain steganalyzer SRNet.

2. According to the diffusion model-based carrier generation method for large-capacity color image steganography according to claim 1, it is characterized by: The encoder part consists of a dry convolution block and four downsampling blocks. Each downsampling block contains two convolution layers and a maximum pooling layer. Each time the feature map passes through a downsampling block, the size changes to half of the previous layer, and the number of channels doubles compared to the previous layer. The outputs of the dry convolution block and the first three downsampling blocks are input into the fusion module to perform feature fusion operations. The decoder contains four upsampling blocks, which perform upsampling through bilinear interpolation operations and connect the feature maps obtained by the fusion module to finally restore the feature maps to the size of the original image.

3. According to claim 1, the carrier generation type color image large-capacity steganography method based on the diffusion model is characterized by: The loss function includes the loss function of the image steganography part and the loss function of the carrier image generation part.

4. According to claim 3, the carrier generation type color image large-capacity steganography method based on the diffusion model is characterized by: The specific steps of the loss function of the image steganography part are: S51: Use mean square error loss to measure the difference between the steganographic image and the carrier image: （1） in: is the steganalysis loss between the stego-image and the carrier image; is the pixel value of the carrier image; is the pixel value of the stego-image; is the number of pixels of the image; S52: In order to narrow the gap between the extracted image and the original secret image, the model is trained using a loss that optimizes the visual similarity between the extracted image and the original embedded secret image. , which is calculated as follows: （2） in: is the pixel value of the carrier image; is the pixel value of the stego-image; is the number of pixels of the image; S53: Based on the structural similarity loss, a loss is designed between the carrier image and the stego image. L STE2 ; The loss between the secret image and the extracted image L EXT2 ; （3） （4） in: and is the mean and variance of the carrier image; and is the mean and variance of the encrypted image; and represents the mean and variance of the secret image; and Indicates the mean and variance of the extracted image; represents the covariance between the carrier image and the encrypted image; represents the covariance between the secret image and the extracted image; In the formula , L is the dynamic range of the image pixel values, K i is a coefficient much smaller than 1; Based on the structural similarity loss function, it not only improves the similarity between the secret image and the carrier image, but also ensures the visual consistency between the extracted image and the secret image. S54: In the steganalysis adversarial network, the steganalyzer discriminates the steganalyzer image and generates a discriminant loss L SAN， Improve the steganalysis resistance of the generated steganalysis image by feeding it back to the generator; （5） in: , is the Softmax output of the steganalyzer XuNet, , is the corresponding truth value; S55: Total loss function of the generator of the image steganography part L S for: （6） The total loss function of the extractor in the image extraction part L E for: （7）。 5. According to claim 4, the carrier generation type color image large-capacity steganography method based on the diffusion model is characterized by: The steps of the loss function for the carrier image generation part are: S61: Using mean square error loss to measure the difference between images can effectively reduce the distance between two images. The calculation formula is defined as: （8） in: is the pixel value of the target image; is the pixel value of the predicted image; is the number of pixels of the image; S62: Total loss function of carrier image generation based on diffusion model Total_DIF as follows: （9） in: It is the weight for adjusting the effect of loss.

Images