High-capacity reversible image steganography method and system based on de-noising probability diffusion model

Abstract

The invention discloses a denoising probability diffusion model-based high-capacity reversible image steganography method, which comprises the following steps of: secret information mapping: converting original information into a binary bit string in a coding manner, and generating a group of noise images according to a 1bpp or 2bpp data mapping rule to obtain noise conversion data; noise image denoising and sampling: denoising and diffusing the noise conversion data iteratively based on an improved denoising diffusion probability model to generate a new sample; according to the method, reverse denoising is performed based on a variational automatic encoder, a denoising diffusion model adopts a forward noise adding process for existing pixel distribution on the basis of VAE, and reverse sampling is opposite to forward diffusion: an original image is extracted from an image degraded into Gaussian distribution; secret information extraction is carried out by using a convolutional neural network; by utilizing the characteristic that the de-noising probability diffusion model has high embedding capacity, more messages can be embedded in pictures with the same size, so that the cost and the times required by transmission are saved, and the resource consumption and the possibility of being discovered are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of information steganography, and in particular relates to a large-capacity reversible image steganography method and system based on a denoising probability diffusion model. Background Art

[0002] Early information hiding methods usually modify the statistical characteristics of the carrier to embed the secret information into the carrier. These methods are divided into spatial domain and transform domain. Typical spatial domain methods include LSB, WOW, HUGO, etc., and transform domain methods include DWT and DCT. However, these methods are easily detected by steganalysis.

[0003] In order to fight against classical steganalysis, carrier-free information hiding has been rapidly developed. This method does not modify the carrier and can effectively resist steganalysis. For carrier-free image steganography, it can be divided into mapping-based image steganography methods and generative-based image steganography methods. In the classical mapping method, Zhou et al. (Zhou Z, Sun H, Harit R, et al. Coverless image steganography without embedding [C] / / Cloud Computing and Security: First International Conference, ICCCS 2015, Nanjing, China, August 13-15, 2015. Revised Selected Papers 1. Springer International Publishing, 2015: 123-132) generate hash values ​​by comparing the average pixel values ​​between each block. On this basis, Zheng et al. (Zheng S, Wang L, Ling B, et al. Coverless information hiding based on robust image hashing [C] / / Intelligent Computing Methodologies: 13th International Conference, ICIC 2017, Liverpool, UK, August 7-10, 2017, Proceedings, Part III13. Springer International Publishing, 2017: 536-547.) used scale-invariant features to improve the robustness of the algorithm, greatly improving the security of the original algorithm. Liu et al. (Q. Liu, X. Xiang, J. Qin, Y. Tan, J. Tan, Y. Luo, Coverless steganography based on image retrieval of DenseNet features and DWT sequence mapping, Knowl.-Based Syst. 192 (2020) 105375.) used the features of discrete wavelet transform and dense convolutional network for sequence mapping.Then Zhang et al. (Zhang X, Peng F, Long M. Robust coverless image steganography based on DCT and LDA topic classification [J]. IEEE Transactions on Multimedia, 2018, 20 (12): 3223-3238) proposed a highly robust algorithm based on discrete cosine transform and latent Dirichlet allocation. With the development of artificial intelligence technology, generative steganography has emerged. Different from the mapping-based steganography method, the classic generative steganography uses the generative adversarial network (GAN). Duan et al. (Duan X, Song H. Coverless information hiding based on generative model [J]. arXiv preprint arXiv: 1802.03528, 2018.) first proposed the GAN-based carrier-free image steganography in 2018. In order to combine the mapping-based and generation-based methods, based on Duan's research, Chen et al. (Chen X, Zhang Z, Qiu A, et al. A novel coverless steganography method based on image selection and StarGAN [J]. IEEE Transactions on Network Science and Engineering, 2020, PP (99): 1-1) proposed a StarGAN-based method, which achieved good results. Later, Peng et al. proposed an approximate method based on generative adversarial networks and gradient descent, which greatly improved the embedding capacity.Hu et al. (Donghui H, Liang W, Wenjie J, et al. A Novel Image Steganography Method via Deep Convolutional Generative Adversarial Networks [J]. IEEE Access, 2018, 638303-38314) proposed a generative steganography method based on deep convolutional generative adversarial network (DCGAN). This method maps secret information into a noise vector, uses a generative adversarial network to generate a natural image, and then uses an extractor to restore the secret information. The ability to resist steganalysis is also considered in the performance of the algorithm. Although Hu's scheme has high extraction accuracy and larger embedding capacity than traditional information hiding methods, the adversarial training and debugging of this scheme is difficult, and the embedding capacity needs to be further improved. Huang Yin et al.

[14] A dual-branch generative adversarial network is proposed to improve the embedding capacity and model accuracy.

[0004] With the further development of steganalysis detection and analysis technology, GAN technology has the following problems: First, the accuracy of information extraction is limited. It is difficult to accurately extract non-decoupled features from images with existing methods. Second, the image generation effect is poor and the diversity is poor. The same features will appear in the generation process. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a large-capacity reversible image steganography method based on an improved denoising probability diffusion model. Aiming at some problems in DDPM, an improved IDDPM is proposed on this basis. After the improved IDDPM undergoes the reverse Markov chain process, random Gaussian noise can be denoised into a natural image. Using IDDPM as a denoising diffusion model is suitable for information hiding in low-resolution images.

[0006] In order to achieve the above-mentioned object, the present invention provides a large-capacity reversible image steganography method based on a denoising probability diffusion model in the first aspect, characterized in that it includes a secret information mapping stage, a noise image denoising sampling stage and a secret information extraction stage;

[0007] Secret information mapping stage: The original information is converted into a binary bit string by encoding, and a set of noise images is generated according to the 1bpp or 2bpp data mapping rules to obtain noise conversion data;

[0008] Noise image denoising sampling stage: based on the improved denoising diffusion probability model, the noise conversion data is iteratively denoised and diffused to generate new samples;

[0009] Based on the variational autoencoder reverse denoising, the denoising diffusion model uses a forward denoising process for the existing pixel distribution on the basis of VAE. The reverse sampling is opposite to the forward diffusion: extracting the original image from the image that has degenerated into a Gaussian distribution;

[0010] In the secret information extraction stage, a convolutional neural network-based extractor is used to extract secret information.

[0011] Furthermore, the improved denoising diffusion probability model is based on variational autoencoder.

[0012]

[0013] Each sample x is generated by the latent variable z, and the latent variable z is generated by the posterior probability network Learned from.

[0014] Furthermore, the denoising diffusion model is based on VAE, given a specific pixel distribution x 0 ~q(x 0 ), define a forward noise adding process, at x 1 On by x T Gaussian noise is added at time t, and the noise variance is β T ∈(0,1), the forward diffusion process is a process in which both the standard deviation and the mean are fixed. The forward noise addition process is as follows:

[0015]

[0016] Solve the above formula to get x t The distribution state of x is reshaped by parameter reshaping t Sampling, by calculating x 0 and β t , we can get any p(x t ):

[0017]

[0018] Among them, α t =1-β t , And β t will grow over time, α t It will become smaller.

[0019] Furthermore, the reverse process is calculated as follows:

[0020]

[0021] The target likelihood function of the diffusion model is:

[0022]

[0023] L hybrid =L simple +λL vlb

[0024] Where λ is a constant, D KL Refers to KL divergence. The improvement of IDDPM is that it uses L hybrid Instead of the original loss function, the algorithm is more suitable for low-resolution images. The goal of diffusion model training is to minimize L hybrid .

[0025] Furthermore, the extractor adopts a convolutional neural network, including a convolution layer, a batch normalization layer and an activation function. There are 9 layers, the first layer is the input layer, the ninth layer is the output layer, a batch normalization layer is provided after the convolution layer, the activation function is provided after each convolution layer and batch normalization, and the Adam optimizer is used to optimize the learning rate.

[0026] Furthermore, the batch normalization layer is placed after the convolutional layer. The extractor uses the batch normalization layer for processing. Specifically, the distribution of each feature is adjusted to a standard normal distribution with a mean of 0 and a variance of 1. For a given batch data X = {x 1 ,x 2 ,...,x m}, m is the batch size, calculate the mean and variance of the current batch, and then perform normalization and translation scaling:

[0027]

[0028] Among them, ε takes a minimum value to prevent the variance from being zero, and γ and β are learnable parameters used to scale and translate the normalized data.

[0029] Furthermore, LeakyReLU introduces a small slope when the input is negative, and the mathematical expression of the activation function is as follows:

[0030]

[0031] Here, α is a constant less than 1.

[0032] In a second aspect, the present invention also provides a large-capacity reversible image steganography system based on a denoising probability diffusion model, including a secret information mapping module, a noise image denoising sampling module and a secret information extraction module;

[0033] The secret information mapping module is used to convert the original information into a binary bit string by encoding, generate a set of noise images according to the 1bpp or 2bpp data mapping rule, and obtain noise conversion data;

[0034] The noise image denoising sampling module iteratively denoises and diffuses the noise conversion data based on the improved denoising diffusion probability model to generate new samples; based on the variational autoencoder reverse denoising, the denoising diffusion model uses a forward denoising process for the existing pixel distribution on the basis of VAE, and the reverse sampling is the opposite of forward diffusion: extracting the original image from the image that has degenerated into a Gaussian distribution;

[0035] The secret information extraction module uses a convolutional neural network-based extractor to extract secret information.

[0036] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the large-capacity reversible image steganography method based on the denoising probability diffusion model described in the present invention.

[0037] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, the large-capacity reversible image steganography method based on a denoising probability diffusion model of the present invention can be implemented.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects: by utilizing the high embedding capacity characteristic of the denoising probability diffusion model, more messages can be embedded in pictures of the same size, saving the overhead and number of times required for transmission, reducing resource consumption and the possibility of being discovered; the method can be applied to training models under a variety of data sets, can be selected in a variety of sizes, and can select appropriate models for use according to different tasks, thereby improving the ability to cope with complex application environments, and the extraction accuracy is higher than that of traditional generative methods. At the same time, the quality of the generated images is better and the richness is more diverse, which can better disguise the transmitted images as natural images and prevent the adversary from discovering the existence of covert transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is the algorithm framework and process of the present invention.

[0040] Figure 2 It is a single-layer neural network structure.

[0041] Figure 3 Schematic diagram of the extractor structure.

[0042] Figure 4To generate a curve graph of the FID value of the secret image as the number of sampling steps changes.

[0043] Figure 5 Generate images for secret information under different datasets and different sampling steps. DETAILED DESCRIPTION

[0044] The method flow of the present invention is as follows Figure 1 As shown, it includes a secret information mapping stage, a noise image denoising sampling stage and a secret information extraction stage; a sender Alice and a receiver Bob are set. When loading the secret information, Alice maps the plaintext data into a noise image according to the prescribed mapping rules, and generates the key Key of the conversion process after fitting the distribution of the noise image to a Gaussian distribution. Then, the noise image is input into the IDDPM, and the Gaussian prediction denoising of the IDDPM is used to gradually transform the noise image into a meaningful normal image. Alice transmits the information to Bob. When extracting the secret information, Bob inputs the denoised image into the extractor to obtain the original noise image with the secret information, and then restores the original noise image using the Key and inversely maps it to obtain the secret information. The Key is a meaningless binary string to any attacker.

[0045] Secret information mapping: In order to express the secret information in the form of vectors, the secret information is first mapped. Whether it is in the form of text or image, it can be converted into a binary bit string by encoding. The noise vector calculated in the neural network, this method randomly takes values ​​between [-1,1]. Therefore, for the bit string, the cloud server can generate a set of noise images according to the rules in Table 2. The two different mapping rules represent the performance advantages and disadvantages of the two mappings. 1bpp gives up higher embedding capacity in exchange for higher extraction accuracy; while 2bpp or even higher embedding capacity embeds more information in the same pixel position at the expense of extraction accuracy.

[0046] Table 1 1bpp data mapping rules

[0047]

[0048] Table 2 2bpp data mapping rules

[0049]

[0050]

[0051] IDDPM denoising process: The improved denoising diffusion probability model IDDPM has high generation ability and sample quality, and is widely used in image generation and text generation. The improved denoising diffusion probability model is based on the idea of ​​diffusion process and denoising autoencoder. It denoises and diffuses the data iteratively to learn the distribution of the data and generate new samples. The model combines probabilistic modeling and deep learning techniques to generate high-fidelity data samples without explicit sampling. It has excellent denoising ability for the noisy images generated in this method.

[0052] The improved denoising probability diffusion model uses a technique for generating models, which is derived from the basic framework of the diffusion model. The diffusion model generates new samples by gradually adding noise to the data and learning the reverse process to remove the noise. This type of model has shown excellent results in the fields of image generation and speech synthesis. The improved denoising probability diffusion model can be used to generate high-quality images for a wide range of artistic creations and product design; in speech generation, the model can also generate natural and fluent speech, which is used in industries such as assistants and entertainment. The model needs to learn how to gradually recover from the noise to the data distribution. This process can be seen as a gradual transition from a simple distribution (such as a Gaussian distribution) to the target data distribution.

[0053] In the improved denoising probability diffusion model, the cosine scheme is used to generate the noise parameter β, which plays a key role in the model training and generation process. The model predicts the linear weight between β and the true variance, and combines the mean square error and KL divergence to optimize the model. The improved denoising probability diffusion model adopts the UNet structure. The UNet structure of IDDPM contains ResBlock and AttentionBlock. ResBlock is used to fuse time information and image features, while AttentionBlock is based on a multi-head self-attention mechanism to capture the relationship between pixels. The model applies the attention mechanism at different resolutions to enhance feature extraction capabilities. Time Embedding: In super-resolution tasks, time t is integrated into the model's prediction through Embedding, which helps the model better understand time series data, thereby improving the quality of generated images or videos.

[0054] The improved IDDPM has a better generation effect for low-resolution images, especially when the number of iterations is small, high-quality natural images can be obtained. IDDPM optimizes the log-likelihood, that is, the cross entropy, while ensuring good FID (Fréchet Inception Distance) and IS (Inception Score) indicators. For the reverse denoising process, its task is to restore the noise vector of the original data in the image, video and other data containing a lot of noise.

[0055] Based on Variational Autoencoder (VAE):

[0056]

[0057] Each sample x is generated by the latent variable z, and the latent variable z is generated by the posterior probability network The goal of VAE is to learn the potential distribution of data so that the model can generate new data samples with similar distribution. The model is trained by maximizing the likelihood of the data and minimizing the KL divergence between the potential representation and the prior distribution to ensure that the learned potential representation can be evenly distributed in the prior distribution. For the denoising diffusion model, based on VAE, given a specific pixel distribution x 0 ~q(x 0 ), define a forward noise adding process, at x 1 On by x T Gaussian noise is added at time t, and the noise variance is β T ∈(0,1). In the process of forward diffusion, it is essentially a process with fixed standard deviation and mean, and a Markov chain without parameter setting. The formula of the forward noise addition process is:

[0058]

[0059] The above formula can be used to solve x t The distribution state of x is then reshaped by parameter reshaping t Sampling. At the same time, by calculating x 0 and β t , we can get any p(x t ):

[0060]

[0061] Among them, α t =1-β t , And β t will grow over time, so α t It will become smaller.

[0062] The reverse sampling process is the opposite of forward diffusion. The reverse process is still a Markov process, which can extract the original image from the image that has degraded to a Gaussian distribution. In the reverse reasoning process, the distribution of data cannot be calculated directly. It is necessary to learn the original image and the degraded image through a neural network, optimize the model, and predict the pixel value in the reverse process to obtain the original image. The formula of the reverse process is expressed as:

[0063]

[0064] In IDDPM, in addition to adapting to small-sized images, the target likelihood of the diffusion model, namely the cross entropy, is also guaranteed. The target likelihood function of the diffusion model is:

[0065]

[0066] L hybrid =L simple +λL vlb (11)

[0067] where λ is set to 0.001, D KL Refers to KL divergence. The improvement of IDDPM is that it uses L hybrid Instead of the original loss function, the algorithm is more suitable for low-resolution images. The goal of diffusion model training is to minimize L hybrid , thereby improving the reverse accuracy of the model.

[0068] The extractor of the present invention adopts a convolutional neural network, including a convolution layer, a batch normalization layer and an activation function. The single-layer neural network structure is as follows: Figure 2 . The Adam optimizer is used to optimize the learning rate. Adam optimization is a commonly used gradient descent optimization algorithm that combines the characteristics of the momentum method and adaptive learning rate. It is widely used in deep learning and can effectively adjust the learning rate and adaptively adjust the learning rate of each parameter during training. The extractor uses a method of keeping the size of the layers unchanged to retain more original information throughout the network. This is particularly important for some tasks that require fine-grained feature information (such as image segmentation, object detection, etc.). It also helps the training of deep networks and reduces the risk of overfitting. It does not require additional steps to process feature maps of inconsistent sizes, which can improve the training speed and inference speed of the network.

[0069] like Figure 3 As shown, the extractor has 9 layers as a whole. Except for the 1st and 9th layers for input and output, batch normalization and activation functions are not required. Figure 2 Connect the structures shown.

[0070] Batch Normalization Layer After the convolutional layer, the extractor uses a batch normalization layer for processing. Its function is to adjust the distribution of each feature to a standard normal distribution with a mean of 0 and a variance of 1. For a given batch data X = {x 1 ,x 2 ,...,x m}, where m is the batch size. Calculate the mean and variance of the current batch:

[0071]

[0072] Perform normalization and translation scaling:

[0073]

[0074] In formula (14), ∈ takes a minimum value to prevent the variance from being zero, and γ and β are learnable parameters used to scale and translate the normalized data. This layer solves the gradient vanishing and gradient exploding problems, accelerates the convergence speed, and improves the generalization ability of the model.

[0075] In the extractor, LeakyReLU is used as the activation function, which is placed after each convolution layer and batch normalization. LeakyReLU is a variant of the Rectified Linear Unit. LeakyReLU introduces a small slope when the input is negative, instead of setting the negative part directly to 0. The mathematical expression of the activation function is as follows:

[0076]

[0077] In formula (16), α is a parameter less than 1, usually with a very small value, such as 0.01. Compared with the traditional ReLu activation function, LeakyRelu can retain certain information in the negative part, and can solve the problem of gradient disappearance when training deep neural networks. At the same time, it also improves the versatility of the model and utilizes nonlinear characteristics to help learn more complex models.

[0078] In order to evaluate the steganographic performance of this scheme, the embedding capacity, training time and image generation quality are tested. The environment configuration used in the experiment is: Intel i5-13400f 2.5GHz CPU, RTX 4060Ti GPU, 32GBDDR4 RAM, professional version of Pycharm with python3.11 desktop computer, using Pytorch as the deep learning framework.

[0079] Embedding capacity refers to the amount or size of secret information that can be hidden in the carrier data. Embedding capacity is usually measured in bits or bytes, indicating the maximum capacity of secret information that can be embedded in the carrier data. For some GAN-based mapping-based steganography, noise is mapped into a vector for generative adversarial generation. In a 64-pixel square image, only 100-dimensional vectors are embedded, and the amount of information in each dimensional vector is between 1 and 3 bits. Its steganographic capacity is 2.44×10 -2 bit / pixel~7.32×10 -2The image size used by Otori et al. in the paper Texture synthesis for mobile data communications is large, but the embedding capacity is not large. Thanks to the good generation effect of the diffusion model, this scheme can achieve an embedding capacity of 1bpp to 3bpp with high extraction accuracy. Therefore, this scheme has the characteristic of high embedding capacity. The comparison between image size and embedding capacity of different schemes is shown in Table 3. Three existing schemes are cited. Control scheme 1 is described in the paper Otori H, Kuriyama S. Texture synthesis for mobile data communications [J]. IEEE Computer Graphics and Applications. 2009, 29 (6): 74-81, control scheme 2 is described in the paper Liu, X., Ma, Z., Ma, J., et al. Image disentanglement autoencoder for steganography without embedding [J]. CVPR 2022, and control scheme 3 is described in the paper Donghui H, Liang W, Wenjie J, et al. A Novel Image Steganography Method via Deep Convolutional Generative Adversarial Networks [J]. IEEE Access, 2018, 638303-38314.

[0080] Table 3 Comparison of image size and embedding capacity of different schemes

[0081]

[0082] The embedding capacity and extraction accuracy were tested. As shown in Table 4, the extraction accuracy of the method of the present invention is much higher than that of the control scheme 2. The best extraction effect is achieved when the embedding capacity is 1bpp and 2bpp. At 4bpp, the algorithm extraction accuracy is close to that of the control scheme 2. At the same time, with the increase of the number of training rounds epoch, the data with the number of training rounds ranging from 10K to 40K were tested. When the sampling step number steps=100, the learning rate is 0.0001, and the batchsize is set to 128, the changes in the extraction accuracy with the increase of the number of training rounds under the embedding rate of 1 to 3bpp are tested. It can be seen from Table 5 that for each embedding rate, the extraction accuracy will increase with the increase of the number of training rounds, and the rate of this increase is gradually slow. When the training epochs is 40K, the 1bpp extraction accuracy reaches 95.65%.

[0083] Table 4 Comparison of extraction accuracy under different embedding capacities

[0084]

[0085] Table 5 Relationship between extraction accuracy and number of training rounds under different embedding capacities

[0086]

[0087] The training time of the model represents the total time it takes for the model to iteratively learn parameters on the training dataset. During the training process, the model will repeatedly adjust the parameters through the dataset so that the model can better fit the data. The length of the training time depends on factors such as the size of the training dataset, the complexity of the model, and the availability of computing resources. As shown in Table 6, the training time per 10K rounds is stable at 40 GPU hours. When 40K rounds are reached, it takes 160.8 hours, about 6.5 days.

[0088] Table 6 Training time

[0089]

[0090] The image generation effect determines the performance of not being detected when transmitted in the channel. Good natural images have high camouflage. Figure 4As shown in the figure, the FID values ​​of the images generated by ImageNet64×64 and CIFAR-10 are tested. The smaller the value, the closer it is to natural images and the less likely it is to be detected. It can be seen that when the sampling step is steps = 100, the FID value is already low. Considering the trade-off between computational cost and image quality, the algorithm selects steps = 100 as the sampling round number. At this time, the generated image quality is better and the required computational cost is relatively small. When the embedding capacity is steps = 100, the impact on FID is small and can be ignored. Figure 5 In the figure, the image generation effects of various datasets at different sampling steps are shown. It can be seen intuitively that under the condition of steps = 100, the generation effects of ImageNet64×64 and CIFAR-10 datasets are better and more diverse.

[0091] Based on the technical concept of the method described in the present invention, a large-capacity reversible image steganography system based on a denoising probability diffusion model can also be provided, including a secret information mapping module, a noise image denoising sampling module and a secret information extraction module;

[0092] The secret information mapping module is used to convert the original information into a binary bit string by encoding, generate a set of noise images according to the 1bpp or 2bpp data mapping rule, and obtain noise conversion data;

[0093] The noise image denoising sampling module iteratively denoises and diffuses the noise conversion data based on the improved denoising diffusion probability model to generate new samples; based on the variational autoencoder reverse denoising, the denoising diffusion model uses a forward denoising process for the existing pixel distribution on the basis of VAE, and the reverse sampling is the opposite of forward diffusion: extracting the original image from the image that has degenerated into a Gaussian distribution;

[0094] The secret information extraction module uses a convolutional neural network-based extractor to extract secret information.

[0095] On the other hand, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the large-capacity reversible image steganography method based on a denoising probability diffusion model described in the present invention can be implemented.

[0096] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the large-capacity reversible image steganography method based on the denoising probability diffusion model described in the present invention.

[0097] The computer device may be a laptop computer, a desktop computer or a workstation.

[0098] The processor may be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or an off-the-shelf field programmable gate array (FPGA).

[0099] The memory of the present invention may be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0100] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drive (SSD) or optical disk, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0101] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A large-capacity reversible image steganography method based on a denoising probability diffusion model, characterized in that: It includes secret information mapping stage, noise image denoising sampling stage and secret information extraction stage; Secret information mapping stage: The original information is converted into a binary bit string by encoding, and a set of noise images is generated according to the 1bpp or 2bpp data mapping rules to obtain noise conversion data; Noise image denoising sampling stage: based on the improved denoising diffusion probability model, the noise conversion data is iteratively denoised and diffused to generate new samples; Based on the variational autoencoder reverse denoising, the denoising diffusion model uses a forward denoising process for the existing pixel distribution on the basis of VAE. The reverse sampling is opposite to the forward diffusion: extracting the original image from the image that has degenerated into a Gaussian distribution; In the secret information extraction stage, a convolutional neural network-based extractor is used to extract secret information.

2. According to claim 1, the large-capacity reversible image steganography method based on the denoising probability diffusion model is characterized in that: The improved denoising diffusion probability model is based on variational autoencoder, Each sample x is generated by the latent variable z, and the latent variable z is generated by the posterior probability network Learned from.

3. The large-capacity reversible image steganography method based on denoising probability diffusion model according to claim 2 is characterized in that: Based on VAE, the denoising diffusion model defines a forward denoising process based on a specific pixel distribution x0~q(x0) by x1. T Gaussian noise is added at time t, and the noise variance is β T ∈(0,1), the forward diffusion process is a process in which both the standard deviation and the mean are fixed. The forward noise addition process is as follows: Solve the above formula to get x t The distribution state of x is reshaped by parameter reshaping t Sampling, by calculating x0 and β t , we can get any p(x t ): Among them, α t =1-β t , And β t will grow over time, α t It will become smaller.

4. The large-capacity reversible image steganography method based on denoising probability diffusion model according to claim 3 is characterized in that: The reverse calculation is as follows: The target likelihood function of the diffusion model is: THE hybrid =L simple +λL vlb Where λ is a constant, D KL Refers to KL divergence. The improvement of IDDPM is to use L hybrid Instead of the original loss function, the algorithm is more suitable for low-resolution images. The goal of diffusion model training is to minimize L hybrid .

5. The large-capacity reversible image steganography method based on denoising probability diffusion model according to claim 1 is characterized in that: The extractor adopts a convolutional neural network, including a convolution layer, a batch normalization layer and an activation function. There are 9 layers, the first layer is the input layer, the ninth layer is the output layer, a batch normalization layer is set after the convolution layer, and the activation function is set after each convolution layer and batch normalization. The Adam optimizer is used to optimize the learning rate.

6. The large-capacity reversible image steganography method based on denoising probability diffusion model according to claim 5 is characterized in that: Batch Normalization Layer After the convolutional layer, the extractor uses a batch normalization layer to process. Specifically, the distribution of each feature is adjusted to a standard normal distribution with a mean of 0 and a variance of 1. For a given batch data X = {x1, x2, ..., x m }, m is the batch size, calculate the mean and variance of the current batch, and then perform normalization and translation scaling: Among them, ε takes a minimum value to prevent the variance from being zero, and γ and β are learnable parameters used to scale and translate the normalized data.

7. The large-capacity reversible image steganography method based on denoising probability diffusion model according to claim 5 is characterized in that: LeakyReLU introduces a small slope when the input is negative. The mathematical expression of the activation function is as follows: Here, α is a constant less than 1.

8. A large-capacity reversible image steganography system based on a denoising probability diffusion model, characterized in that: It includes a secret information mapping module, a noise image denoising sampling module and a secret information extraction module; The secret information mapping module is used to convert the original information into a binary bit string by encoding, generate a set of noise images according to the 1bpp or 2bpp data mapping rule, and obtain noise conversion data; The noise image denoising sampling module iteratively denoises and diffuses the noise conversion data based on the improved denoising diffusion probability model to generate new samples; based on the variational autoencoder reverse denoising, the denoising diffusion model uses a forward denoising process for the existing pixel distribution on the basis of VAE, and the reverse sampling is the opposite of forward diffusion: extracting the original image from the image that has degenerated into a Gaussian distribution; The secret information extraction module uses a convolutional neural network-based extractor to extract secret information.

9. A computer device, characterized in that: It includes a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can implement the large-capacity reversible image steganography method based on the denoising probability diffusion model as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, the large-capacity reversible image steganography method based on a denoising probability diffusion model as described in any one of claims 1 to 6 can be implemented.

Images