Anti-noise high-capacity robust image steganography method
Through cascading network structure and loss function training, real-world noise interference is simulated, and the stability and security problems of image steganography method under complex channels are solved, achieving high-fidelity recovery of large-capacity secret information and the robustness of steganography system.
Patent Information
- Application Number
- CN202510568270.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-15
AI Technical Summary
Existing image steganography methods are difficult to achieve a good balance between robustness, high capacity and anti-steganography analysis. Especially under complex channels such as social platforms, the stability and security of the steganography system are fragile, making it difficult to maintain high-fidelity recovery when large-capacity secret information is embedded.
The cascading structure of reversible current mapping network, latent variable steganography network, noise simulation network and deep recovery network is adopted. Through discrete wavelet transformation and loss function training, real-world noise interference is simulated and secret images are generated and restored.
In complex noise environments, the recovery quality and stability of secret images are improved, the anti-interference ability of the steganography system is enhanced, multi-image steganography and social network communication are supported, and the application boundaries of steganography technology are improved.
Smart Images
Figure CN120495057A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of multimedia content security and image steganography technology, and in particular to a noise-resistant, large-capacity, robust image steganography method. Background Art
[0002] Image steganography (image-within-image) aims to embed a full-sized, color secret image into a carrier image of the same size, losslessly or with minimal loss, to generate a secret-carrying image. This process requires not only that the secret image be imperceptible to ensure concealment, but also that the secret image be recovered with the highest possible fidelity during the information extraction phase. However, because the full-sized secret image significantly exceeds the information hiding capacity of the carrier image, embedding such large amounts of secret information severely degrades the visual quality of the stegoimage, posing a serious threat to the security of the stegoimage. Existing steganography methods struggle to achieve secure hiding of full-sized secret images. Furthermore, in lossy channels such as those found in social networks (e.g., image compression, screen capture, and channel noise), the stegoimage is inevitably affected by the platform's lossy processing. The embedding of large amounts of secret information can cause even minor noise interference to result in significant information loss or even severe distortion in the extracted secret image. Consequently, existing steganography methods struggle to strike a good balance between robustness, high capacity, and resistance to steganalysis. This is particularly vulnerable to the social noise interference of complex channels, such as those found in social platforms. Summary of the Invention
[0003] The purpose of the present invention is to provide a noise-resistant, large-capacity, robust image steganography method that improves the stability and security of the steganography system and improves the quality of the restored image.
[0004] The purpose of the present invention can be achieved by the following technical solutions:
[0005] A noise-resistant, large-capacity, robust image steganography method, comprising the following steps:
[0006] (a) The sender obtains the secret image and the carrier image in the image hiding process. The secret image and the carrier image are subjected to discrete wavelet transform to obtain the high-frequency and low-frequency components of the secret image and the carrier image respectively.
[0007] (b) Constructing a reversible flow mapping network, mapping the high- and low-frequency components of the secret image to latent variables based on the reversible flow mapping network, inputting the latent variables and the high- and low-frequency components of the carrier image into the latent variable steganalysis network, outputting the high- and low-frequency components of the secret image, and obtaining the secret image through inverse discrete wavelet transform;
[0008] (c) Constructing a noise simulation network, applying different degrees of common distortion and noise output by the noise simulation network to the encrypted image to obtain a distorted encrypted image;
[0009] (d) The distorted secret image is processed by the deep recovery network to generate the recovered secret image;
[0010] (e) Based on the recovered secret image and the secret image of (a), a loss function is constructed to train the reversible flow mapping network, the latent variable steganography network, the noise simulation network and the deep recovery network to obtain the image steganography model and the image recovery model. The image steganography model and the image recovery model obtain the actual secret image and the carrier image for steganography and recovery.
[0011] Furthermore, the specific steps of mapping the high- and low-frequency components of the secret image to the latent variables based on the reversible flow mapping network are as follows:
[0012] The reversible flow mapping network includes a reversible 1×1 convolutional layer and an affine coupling layer. The alternating coupling process in the affine coupling layer consists of three layers of residual convolution blocks. The high-frequency and low-frequency components of the secret image pass through the reversible 1×1 convolutional layer and the affine coupling layer in sequence to generate latent variables, which include the low-frequency and high-frequency components of the latent variables.
[0013] Furthermore, the latent variables are:
[0014] y1=x1·exp(α(s(x2)))+s(x2)
[0015] y2=x2·exp(α(s(y1)))+s(y1)
[0016] Among them, y1 and y2 are the output ends, that is, the low-frequency and high-frequency components of the latent variables, x1 and x2 are the input ends, that is, the high-frequency and low-frequency components of the secret image, α() is the activation function, and s() is the residual convolution block.
[0017] Furthermore, the high- and low-frequency components of the latent variable and the carrier image are input into the latent variable steganographic network, and the specific steps of outputting the high- and low-frequency components of the carrier image are as follows:
[0018] A latent variable steganography network consisting of a U-Net with four layers of downsampling and upsampling structures is constructed. The latent variables and the high and low frequency components of the carrier image are input into the latent variable steganography network. During the downsampling process, the convolution layer is used to reduce the spatial size of the input by half, while the channel information is increased by 2 times. During the upsampling process, the transposed convolution operation is used to enlarge the spatial size and restore the channel size. Skip connections are used between the downsampling and upsampling structures.
[0019] Further, common distortions and noise include a noise floor and a real-world screen shot noise floor.
[0020] Furthermore, the screen capture noise layer includes perspective distortion, moiré, color distortion, device noise, and JPEG compression. When perspective distortion is applied to the encrypted image, a random offset is applied to each of the four vertices of the encrypted image.
[0021] When moiré is applied to a dense image, two sine waves are generated to simulate the moiré effect and applied to each channel of the dense image;
[0022] Randomly adjust the saturation, brightness and contrast of the encrypted image to simulate color distortion;
[0023] Apply Gaussian noise to the encrypted image to simulate device noise;
[0024] Emulate JPEG compression via stochastically differentiable JPEG compression.
[0025] Furthermore, the specific steps of processing the distorted secret image through the deep recovery network to generate the recovered secret image are as follows:
[0026] The distorted secret image is input into the deep recovery network and then obtained through discrete wavelet transform. The high and low frequency components of the distorted secret image are input into the same latent variable steganalysis network as in (b), and then pass through the same reversible flow mapping network as in (b) to generate the recovered secret image through inverse discrete wavelet transform.
[0027] Furthermore, the loss function includes a training recovery process loss function and a training steganography process loss function.
[0028] Furthermore, the loss function of the training recovery process is:
[0029]
[0030] Where N represents the number of training samples, represent the secret image and the secret image recovered from the secret image, respectively.
[0031] Furthermore, the loss function of the steganography training process is:
[0032]
[0033] Where N represents the number of training samples, MSE represents the mean square error, They represent the carrier image and the generated encrypted image respectively.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] The present invention proposes a noise-resistant, large-capacity, robust image steganography method. This method constructs a new, highly robust steganography framework by innovatively designing a universal noise generation simulation network. The present invention can realistically simulate the distortion caused by real-world noise on images in complex distortion environments (such as social network noise, screen capture, etc.), thereby enhancing the adaptability and anti-interference ability of the steganography system to various types of channel noise, effectively improving the recovery quality of the secret image, and ensuring the stable extraction of the steganographic information. At the same time, the framework of the present invention has good scalability and supports multi-image steganography through a cascade structure, making it adaptable to more complex application scenarios, such as multi-layer information hiding, social network steganographic communication, and digital copyright protection, further expanding the application boundaries of steganography technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0037] Figure 2 Schematic diagram of the secret image and the recovered secret image under the JPEG Q=80 distortion condition of the present invention;
[0038] Figure 3 Schematic diagram of recovering a secret image in a real screen shooting scenario according to the present invention. DETAILED DESCRIPTION
[0039] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0040] One of the core challenges facing the current field of image steganography is how to design a new steganographic framework that combines adaptability, information consistency, robustness, and security, while balancing high capacity, noise resistance, and steganalysis security. This paper proposes a noise-resistant, high-capacity, robust image steganography method, which includes the following steps:
[0041] (a) The sender obtains the secret image and the carrier image in the image hiding process. The secret image and the carrier image are subjected to discrete wavelet transform to obtain the high-frequency and low-frequency components of the secret image and the carrier image respectively.
[0042] (b) Constructing a reversible flow mapping network, mapping the high- and low-frequency components of the secret image to latent variables based on the reversible flow mapping network, inputting the latent variables and the high- and low-frequency components of the carrier image into the latent variable steganalysis network, outputting the high- and low-frequency components of the secret image, and obtaining the secret image through inverse discrete wavelet transform;
[0043] (c) Constructing a noise simulation network, applying different degrees of common distortion and noise output by the noise simulation network to the encrypted image to obtain a distorted encrypted image;
[0044] (d) The distorted secret image is processed by the deep recovery network to generate the recovered secret image;
[0045] (e) Based on the recovered secret image and the secret image of (a), a loss function is constructed to train the reversible flow mapping network, the latent variable steganography network, the noise simulation network and the deep recovery network to obtain the image steganography model and the image recovery model. The image steganography model and the image recovery model obtain the actual secret image and the carrier image for steganography and recovery.
[0046] In (a), discrete wavelet transform (DWT) is used to input the wavelet subbands of the secret image and the carrier image in the hiding process, and the wavelet subbands of the distorted carrier image in the recovery process;
[0047] In (b), the secret image and the carrier image are fused through a deep hidden network to generate a secret image;
[0048] In (c), a noise simulation network is constructed to apply different degrees of common distortion and noise to the encrypted image to obtain a distorted encrypted image, thereby simulating the real impact of the social network environment on the encrypted image.
[0049] The distorted secret image in (d) is used as input and processed by the deep recovery network to generate the recovered secret image;
[0050] In (e), a loss function is constructed to train the hidden process, noise simulation network and recovery process.
[0051] The specific process of step (a) is:
[0052] (a1) Decompose the image into low-frequency components and high-frequency components through DWT, and extract the features of the image at different frequencies.
[0053] The specific process of step (b) is:
[0054] (b1) Construct a reversible flow mapping network to map the low-frequency and high-frequency components of the secret image into latent variables;
[0055] (b2) Input the latent variables and carrier images into the latent steganalysis network to achieve deep hiding;
[0056] Furthermore, as a preferred solution, the specific process of step (b1) is:
[0057] (b11) Construct a reversible 1×1 convolutional layer for channel permutation to achieve feature enhancement and accelerate network convergence;
[0058] (b12) Construct an affine coupling layer based on the additive coupling transformation and the multiplicative coupling transformation. The forward process is:
[0059] y1=x1·exp(α(s(x2)))+s(x2)
[0060] y2=x2·exp(α(s(y1)))+s(y1)
[0061] Among them, x1 and x2 are input terminals (representing the low-frequency and high-frequency components of the secret image), y1 and y2 are output terminals (representing the results after coupling mapping), α() is the activation function, and s() is a nonlinear function constructed by the neural network;
[0062] The reverse process is:
[0063] x2=(y2-s(y1))÷exp(α(s(y1)))
[0064] x1=(y1-s(x2))÷exp(α(s(x2)))
[0065] Among them, y1 and y2 are input ends (representing the low-frequency component and high-frequency component of the potential feature), x1 and x2 are output ends (representing the results after coupling mapping), α() is the activation function, and s() is a nonlinear function constructed by the neural network;
[0066] (b13) Construct a residual convolution block s() consisting of convolution operations;
[0067] Furthermore, as a preferred solution, the specific process of step (b11) is:
[0068] (b111) The low-frequency components and high-frequency components of the secret image are superimposed channel by channel to form a feature map of size (h, w, c), where h, w, and c represent the length and width of the feature map and the number of channels, respectively;
[0069] (b112) Generate a rotation matrix using LR decomposition as a custom weight matrix W;
[0070] (b113) Using a 1×1 filter and the weight matrix W, a convolution operation with a step size of 1 is performed on the feature map to obtain a mixed feature after channel permutation.
[0071] The specific process of step (b13) is as follows:
[0072] (b131) Construct a residual convolution block consisting of three convolutional layers. The results of the first two convolutional layers are activated using leakyReLU. The weights of the third convolutional layer are initialized to 0 and its result is directly output.
[0073] (b132) The first two convolution operations are performed sequentially, and the results are residually connected to the original input and input to the third convolution layer.
[0074] The specific process of step (b2) is:
[0075] (b21) Construct a U-Net style hidden network consisting of four layers of downsampling and upsampling structures;
[0076] (b22) During the downsampling process, a convolutional layer with a stride of 2 is used to reduce the spatial size of the input feature map by half, while the channel information is increased by 2 times, thereby extracting higher-level features.
[0077] (b23) During the upsampling process, a transposed convolution operation is used to enlarge the spatial size and restore the channel size.
[0078] (b24) Between upsampling and downsampling, skip connections are used to allow the high-resolution features of the upper layer to assist in recovering the details of the image.
[0079] The specific process of step (c) is as follows:
[0080] (c1) Construct a noise layer, including JPEG compression, Gaussian noise, Gaussian blur, mean filtering, sharpening, and salt and pepper noise. The noise can be combined and adjusted at will to fully adapt to the interference of real social network environments on images.
[0081] (c2) Construct a real-world screen shot noise layer consisting of perspective distortion, moiré, color distortion, device noise, and JPEG compression.
[0082] The specific process of step (c2) is:
[0083] (c21) Perspective distortion simulates the angular bias of a camera when capturing images from various angles, resulting in perspective distortion in the resulting image. Perspective distortion is achieved by applying a random offset (~U[-8,8]) to each of the four vertices of the original encrypted image.
[0084] (c22) Shooting an image on a screen often triggers a moiré effect. This moiré effect is simulated by generating two sine waves and applying them to each channel of the input image.
[0085] (c23) Ambient light, screen color gamut, and camera exposure settings can cause significant color variations in captured images. We randomly adjust the saturation, brightness (~U[0.7,1.3]), and contrast (~U[0.9,1.1]) of the captured images to simulate color variations.
[0086] (c24) Gaussian noise (σ=0.2) is applied to the captured image to simulate the Gaussian noise generated in the camera module.
[0087] (c25) Camera images are usually stored in lossy formats such as JPEG. This process is simulated by using stochastically differentiable JPEG compression (Q~U[65,95]).
[0088] The specific process of step (e) is:
[0089] (e1) Loss function of the steganographic training process:
[0090]
[0091] Where N represents the number of training samples, MSE represents the mean square error, Represent the carrier image and the generated encrypted image respectively;
[0092] (e2) Loss function of the training recovery process
[0093]
[0094] in, They represent the secret image and the secret image finally recovered from the secret image.
[0095] Beneficial effects of the present invention: The present invention proposes a noise-resistant, large-capacity, robust image steganography method. This method constructs a new, highly robust steganography framework by innovatively designing a universal noise generation simulation network. The present invention can realistically simulate the distortion of images caused by real-world noise in complex distortion environments (such as social network noise, screen shooting, etc.), thereby enhancing the adaptability and anti-interference ability of the steganography system to various types of channel noise, effectively improving the recovery quality of secret images, and ensuring the stable extraction of steganographic information. At the same time, the framework of the present invention has good scalability and supports multi-image steganography through a cascade structure, enabling it to adapt to more complex application scenarios, such as multi-layer information hiding, social network steganographic communication, and digital copyright protection, further expanding the application boundaries of steganography technology.
[0096] Figure 1 Flowchart of the present invention.
[0097] The residual convolution block network structure is shown in Table 1.
[0098] Table 1: Residual convolution block network structure table
[0099] Layer Input size kernel size Output size Input 12′(128′128) / 12′(128′128) Conv1 12′(128′128) 64′(3′3) 64′(128′128) Conv2 64′(128′128) 64′(3′3) 64′(128′128) Conv3 (64+12)′(128′128) 12′(3′3) 12′(128′128)
[0100] The network consists of N (N=3) convolutional layers. Each convolutional layer consists of a convolution layer with a convolution kernel size of 3×3 and a stride of 2, and a Leaky ReLU activation function layer. The first two layers have 64 convolution kernels to enrich image features. The third convolution layer uses a residual connection to superimpose the features output by the second convolution layer and the input. The fifth layer has 12 convolution kernels to extract and output image features.
[0101] The hidden network structure based on the autoencoder is shown in Table 2. The network consists of L (L = 4) downsampling and upsampling layers. After processing in the first layer, the input layer is expanded to 64 channels, while maintaining the original resolution. The second through fourth layers successively double the number of feature channels while halving the resolution, using skip connections to preserve feature information. This U-Net-like structure utilizes skip connections to fuse features across different scales, enabling efficient encoding of hidden information at multiple scales, thereby enhancing information hiding capabilities.
[0102] Table 2: Hidden network structure table
[0103] Layer Input size Skip Connection Output size Input 24′(128′128) / 64′(128′128) Layer 1 64′(128′128) 64′(128′128) 128′(64′64) Layer 2 128′(64′64) 128′(64′64) 256′(32′32) Layer 3 256′(32′32) 256′(32′32) 512′(16′16) Layer 4 512′(16′16) 512′(16′16) 512′(16′16)
[0104] The DIV2K training dataset (800 2K resolution images) is used for training and the DIV2K test dataset (100 images), the COCO test dataset (5000 randomly selected images), and the ImageNet test dataset (10,000 randomly selected images) are used for testing. For DIV2K, the resolution of the generated test images is 1024×1024, while for COCO and ImageNet, the resolution is 256×256.
[0105] The experimental settings are as follows: Adam (β1 = 0.5, β2 = 0.999) optimizer is used, and the learning rate is set to 1×10 -4.5 , weight_decay is set to 0.98 and updated every 20 rounds; the total training rounds are 3000 times, the batch-size is set to 2, corresponding to the carrier image and the secret image, and the patch-size is set to 256×256; according to the above settings, the noise-resistant large-capacity robust image steganography model optimizes the loss function through continuous learning and training, so that the forward network can achieve image hiding, and the reverse network can achieve image restoration; the final model parameters are obtained when the parameters in the model remain stable or reach the given maximum number of iterations. Figure 2 and Figure 3The visual presentation of this method is demonstrated in turn. To test the robustness of this method, this example captures classified images from different screen angles and extracts the secret image from them. To test the steganographic security of this method, this example uses the deep learning-based model ZhuNet to test the number of training samples that can be leaked to allow ZhuNet to distinguish between the carrier image and the secret image. The accuracy of extracting the secret image from classified images captured from different screen angles in real-world conditions is shown in Table 3.
[0106] Table 3: Accuracy of extracting secret images from secret images acquired from different screen angles in the real world.
[0107] Screen shooting angle PSNR SSIM 30° left 19.67 0.580 15° left 20.59 0.599 0° 21.37 0.617 15° right 19.42 0.581 30° right 18.88 0.569
[0108] The method of the present invention can accurately extract secret images with good visual quality from secret images acquired from different shooting angles in the real world, while retaining rich semantic information and texture details, fully demonstrating its excellent robustness.
[0109] The results of using ZhuNet to test "how many training samples need to be leaked for ZhuNet to detect anomalies" are shown in Table 4:
[0110] Table 4: Using ZhuNet to test “how many training samples need to be leaked for ZhuNet to detect anomalies”.
[0111]
[0112]
[0113] The method of the present invention is compared with the DeepMIH method, LiDiNet method and StegFormer method based on deep learning. The security performance improvement is obvious, which shows that the method of the present invention has strong anti-steganalysis ability.
[0114] A noise-resistant, large-capacity robust image steganography method is characterized by comprising: utilizing a noise network to simulate noise in real social scenarios to implement a deep steganograph network; utilizing discrete wavelet transform (DWT) to input wavelet subbands of a secret image and a carrier image during the hiding process, and wavelet subbands of a distorted secret image during the recovery process; utilizing a deep hiding network to fuse the secret image and the carrier image to generate a secret image, thereby hiding the secret image; constructing a noise simulation network to apply varying degrees of common distortion and noise to the secret image to obtain a distorted secret image, thereby simulating the real-world effects of a social network environment on the secret image; utilizing a deep recovery network to extract the secret image from the distorted secret image, thereby restoring the secret image; constructing a loss function to train the hiding process, the noise simulation network, and the recovery process; constructing a noise network model by learning complex noise in the real world, thereby enabling high-quality recovery of the secret image under different noise interference scenarios; and greatly improving the invisibility, security, and robustness of large-capacity image steganography, and having the potential to be extended to multi-image steganography.
[0115] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.
Claims
1. A noise-resistant, large-capacity, robust image steganography method, characterized in that: The method comprises the following steps: (a) The sender obtains the secret image and the carrier image in the image hiding process. The secret image and the carrier image are subjected to discrete wavelet transform to obtain the high-frequency and low-frequency components of the secret image and the carrier image respectively. (b) Constructing a reversible flow mapping network, mapping the high- and low-frequency components of the secret image to latent variables based on the reversible flow mapping network, inputting the latent variables and the high- and low-frequency components of the carrier image into the latent variable steganalysis network, outputting the high- and low-frequency components of the secret image, and obtaining the secret image through inverse discrete wavelet transform; (c) Constructing a noise simulation network, applying different degrees of common distortion and noise output by the noise simulation network to the encrypted image to obtain a distorted encrypted image; (d) The distorted secret image is processed by the deep recovery network to generate the recovered secret image; (e) Based on the recovered secret image and the secret image of (a), a loss function is constructed to train the reversible flow mapping network, the latent variable steganography network, the noise simulation network and the deep recovery network to obtain the image steganography model and the image recovery model. The image steganography model and the image recovery model obtain the actual secret image and the carrier image for steganography and recovery.
2. A noise-resistant, large-capacity, robust image steganography method according to claim 1, characterized in that: The specific steps of mapping the high- and low-frequency components of the secret image to latent variables based on the reversible flow mapping network are: The reversible flow mapping network includes a reversible 1×1 convolutional layer and an affine coupling layer. The alternating coupling process in the affine coupling layer consists of three layers of residual convolution blocks. The high-frequency and low-frequency components of the secret image pass through the reversible 1×1 convolutional layer and the affine coupling layer in sequence to generate latent variables. The latent variables include the low-frequency and high-frequency components of the latent variables. In the three convolutional layers of the residual convolution block, the results of the first two convolutional layers are activated using leaky ReLU, and the weight of the third convolutional layer is initialized to 0. Its result is directly used as the output of the third convolutional layer.
3. A noise-resistant, large-capacity, robust image steganography method according to claim 2, characterized in that: The latent variables are: y1=x1·exp(α(s(x2)))+s(x2) y2=x2·exp(α(s(y1)))+s(y1) Among them, y1 and y2 are the output ends, that is, the low-frequency and high-frequency components of the latent variables, x1 and x2 are the input ends, that is, the high-frequency and low-frequency components of the secret image, α() is the activation function, and s() is the residual convolution block.
4. A noise-resistant, large-capacity, robust image steganography method according to claim 1, characterized in that: The specific steps of inputting the latent variable and the high and low frequency components of the carrier image into the latent variable steganalysis network and outputting the high and low frequency components of the carrier image are as follows: A latent variable steganography network consisting of a U-Net with four layers of downsampling and upsampling structures is constructed. The latent variables and the high and low frequency components of the carrier image are input into the latent variable steganography network. During the downsampling process, the convolution layer is used to reduce the spatial size of the input by half, while the channel information is increased by 2 times. During the upsampling process, the transposed convolution operation is used to enlarge the spatial size and restore the channel size. Skip connections are used between the downsampling and upsampling structures.
5. The noise-resistant, large-capacity, robust image steganography method according to claim 1, characterized in that: Common artifacts and noise include noise floors and real-world screen shot noise floors.
6. A noise-resistant, large-capacity, robust image steganography method according to claim 5, characterized in that: The screen capture noise layer includes perspective distortion, moiré, color distortion, device noise, and JPEG compression. When perspective distortion is applied to the encrypted image, a random offset is applied to each of the four vertices of the encrypted image. When moiré is applied to a dense image, two sine waves are generated to simulate the moiré effect and applied to each channel of the dense image; Randomly adjust the saturation, brightness and contrast of the encrypted image to simulate color distortion; Apply Gaussian noise to the encrypted image to simulate device noise; Emulate JPEG compression via stochastically differentiable JPEG compression.
7. A noise-resistant, large-capacity, robust image steganography method according to claim 1, characterized in that: The specific steps of processing the distorted secret image through the deep recovery network to generate the recovered secret image are as follows: The distorted secret image is input into the deep recovery network and then obtained through discrete wavelet transform. The high and low frequency components of the distorted secret image are input into the same latent variable steganalysis network as in (b), and then pass through the same reversible flow mapping network as in (b) to generate the recovered secret image through inverse discrete wavelet transform.
8. The noise-resistant, large-capacity, robust image steganography method according to claim 1, characterized in that: The loss function includes the loss function of the training recovery process and the loss function of the training steganography process.
9. A noise-resistant, large-capacity, robust image steganography method according to claim 8, characterized in that: The loss function of the training recovery process is: Where N represents the number of training samples, represent the secret image and the secret image recovered from the secret image, respectively.
10. A noise-resistant, large-capacity, robust image steganography method according to claim 9, characterized in that: The loss function of the steganography training process is: Where N represents the number of training samples, MSE represents the mean square error, They represent the carrier image and the generated encrypted image respectively.