Massive image steganography method, device and equipment based on reversible neural network

Abstract

The application relates to a large-scale image steganography method, device and computer equipment based on a reversible neural network. The method comprises the following steps: under-sampling secret images to corresponding multiples through a reversible image zooming module; meanwhile, a carrier image is divided into non-overlapping blocks with the same size as the under-sampled secret images; and the generated under-sampled secret images and the carrier image are input into a reversible image hiding module to generate a secret-containing image. The application introduces a reversible image zooming transformation, fully utilizes the bidirectionality of the reversible neural network, overcomes the mutual interference of information and colors among the secret images, and greatly improves the number of image steganography and the quality of recovered decryption images.

Description

Technical Field

[0001] This application relates to the field of image steganography, and in particular to a large-scale image steganography method, apparatus, and computer device based on a reversible neural network. Background Technology

[0002] With the continuous development of deep learning, researchers have proposed spatial domain image steganography methods based on deep learning, such as encoder-decoder network structures based on deep convolutional neural networks. The encoder encodes both the carrier image and the secret image to generate a coded image, and the decoder reconstructs the decrypted image from the coded image. During training, the encoder and decoder are typically coupled and updated alternately, currently capable of hiding and recovering 2-3 secret images. Unlike traditional deep convolutional encoder-decoder neural networks, reversible neural networks share parameters in both forward and backward propagation processes, meaning that image hiding and recovery are considered a bidirectional reversible process. Given the output of one layer of the neural network, the outputs of other layers can be obtained through lossless computation, significantly reducing the amount of parameters required for training and storage. This approach has been increasingly applied to image steganography, further improving the hiding capacity of the carrier image and the quality of the decrypted image reconstruction. Simultaneously, image steganography methods based on reversible neural networks offer a more stable training process and faster convergence speed. Specifically, existing high-capacity image steganography techniques based on reversible neural networks can be broadly divided into two approaches: one is to hide secret images sequentially into a hidden image using a deep hidden network; the other is to concatenate all secret images as a whole as input and then use a deep hidden network to hide them into a carrier image.

[0003] Image steganography methods based on reversible neural networks also require the carrier image and the secret image to be the same size. Therefore, as the number of secret images increases, the information of individual pixels accumulates, leading to mutual interference between the color channels of the secret images. Secret images hidden earlier experience more interference, making the reconstruction of the decrypted image more difficult. Simultaneously, the perturbations added to the hidden image become more noticeable, easily detected by steganalysis algorithms. Although image steganography methods based on reversible neural networks have successfully increased the number of hidden images to five, these problems limit further improvements in the capacity of the carrier image and the quality of the reconstructed decrypted image. Furthermore, in traditional steganalysis methods based on convolutional neural networks and sequential hiding methods based on reversible neural networks, each secret image corresponds to an independent hiding and recovery network, thus facing the problem that the training parameters increase linearly with the number of secret images.

[0004] Therefore, existing technologies suffer from poor adaptability. Summary of the Invention

[0005] Therefore, it is necessary to provide a large-scale image steganography method, apparatus, computer equipment, and storage medium based on reversible neural networks that can effectively improve the capacity of image hiding on the carrier and the quality of image recovery after decryption, in order to address the above-mentioned technical problems.

[0006] A large-scale image steganography method based on a reversible neural network, the method comprising:

[0007] Acquire a carrier image and N secret images to be hidden, and input the carrier image and the secret images into a pre-designed image steganography network model; the image steganography network model includes a reversible image scaling module and a reversible image hiding module;

[0008] The secret image is downsampled using the reversible image scaling module to obtain N scaled-down secret images.

[0009] The carrier image is divided into N non-overlapping blocks, and the reversible image hiding module sequentially hides the reduced secret image into the corresponding non-overlapping blocks to obtain N hidden image blocks; wherein, the size of the reduced secret image is the same as the size of the non-overlapping block;

[0010] The N dense image blocks are sequentially merged to obtain the predicted dense image;

[0011] The image steganalysis network model is trained based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model.

[0012] Image steganography is performed using the trained image steganography network model to obtain a cryptic output image.

[0013] In one embodiment, the method further includes: acquiring a encrypted image to be decrypted;

[0014] The encrypted image to be decrypted is decrypted through the reverse process of the reversible image hiding module to obtain N initially decrypted images;

[0015] The N initial decrypted images are upsampled using the reversible image scaling module to obtain N decrypted images; the image resolution of the decrypted images is the same as that of the original secret images.

[0016] In one embodiment, the method further includes: downsampling the secret image through the convolutional layer in the decomposition block of the reversible image scaling module to obtain low-frequency information and high-frequency information generated by the decomposition block;

[0017] The reversible block in the reversible image scaling module exchanges information based on the low-frequency and high-frequency information generated by the decomposition block to obtain N scaled-down secret images and a variable unrelated to one image.

[0018] In one embodiment, the method further includes: exchanging information between the initially decrypted image and a randomly sampled variable through a reversible block in the reversible image scaling module to obtain low-frequency and high-frequency information output by the reverse process of the reversible block;

[0019] The synthesis block in the reversible image scaling module upsamples the low-frequency and high-frequency information output from the reverse process of the reversible block to obtain N restored decrypted images.

[0020] In one embodiment, the method further includes: decomposing the reduced secret image and the non-overlapping block into low-frequency information and high-frequency information by using a discrete wavelet transform block in the reversible image hiding module;

[0021] The reversible block in the reversible image hiding module exchanges information based on the low-frequency and high-frequency information obtained from the discrete wavelet transform block decomposition, generating N dense image blocks and a variable unrelated to the image.

[0022] In one embodiment, the method further includes: exchanging information between the encrypted image to be decrypted and a randomly sampled variable through the reversible block in the reversible image hiding module to obtain low-frequency information and high-frequency information output by the reversible block;

[0023] The low-frequency and high-frequency information output by the inverse discrete wavelet transform block in the reversible image hiding module is reconstructed to obtain N initially decrypted images.

[0024] In one embodiment, the forward process of the i-th reversible block in the reversible image scaling module and the reversible image hiding module is represented as follows:

[0025]

[0026]

[0027] The reverse process of the i-th reversible block is represented as:

[0028]

[0029]

[0030] in, and These represent the low-frequency and high-frequency information generated by the i-th reversible block, respectively. ρ(·) and ψ(·) both represent dense networks.

[0031] In one embodiment, the method further includes: iteratively training the reversible image scaling module and the reversible image hiding module respectively based on the predicted dense image and a preset loss function, and then jointly training the reversible image scaling module and the reversible image hiding module iteratively; wherein the preset loss function is:

[0032]

[0033] in, The secret image used to guide downsampling contains sufficient information, including the recovery loss function, the downsampling loss function, and the reconstruction loss function; This is the steganalysis loss function, used to constrain the difference between the output cryptic image and the original carrier image to be as small as possible.

[0034] A large-scale image steganography device based on a reversible neural network, the device comprising:

[0035] An image input module is used to acquire a carrier image and N secret images to be hidden, and input the carrier image and the secret images into a pre-designed image steganography network model; the image steganography network model includes a reversible image scaling module and a reversible image hiding module;

[0036] The downsampling module is used to downsample the secret image through the reversible image scaling module to obtain N scaled-down secret images;

[0037] An image hiding module is used to divide the carrier image into N non-overlapping blocks, and to hide the reduced secret image sequentially into the corresponding non-overlapping blocks through the reversible image hiding module, thereby obtaining N hidden image blocks; wherein, the size of the reduced secret image is the same as the size of the non-overlapping block;

[0038] The model prediction module is used to sequentially merge the N dense image blocks to obtain the predicted dense image;

[0039] The model training module is used to train the image steganalysis network model based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model.

[0040] The model application module is used to perform image steganography using the trained image steganography network model to obtain the output steganographic image.

[0041] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0042] Acquire a carrier image and N secret images to be hidden, and input the carrier image and the secret images into a pre-designed image steganography network model; the image steganography network model includes a reversible image scaling module and a reversible image hiding module;

[0043] The secret image is downsampled using the reversible image scaling module to obtain N scaled-down secret images.

[0044] The carrier image is divided into N non-overlapping blocks, and the reversible image hiding module sequentially hides the reduced secret image into the corresponding non-overlapping blocks to obtain N hidden image blocks; wherein, the size of the reduced secret image is the same as the size of the non-overlapping block;

[0045] The N dense image blocks are sequentially merged to obtain the predicted dense image;

[0046] The image steganalysis network model is trained based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model.

[0047] Image steganography is performed using the trained image steganography network model to obtain a cryptic output image.

[0048] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0049] Acquire a carrier image and N secret images to be hidden, and input the carrier image and the secret images into a pre-designed image steganography network model; the image steganography network model includes a reversible image scaling module and a reversible image hiding module;

[0050] The secret image is downsampled using the reversible image scaling module to obtain N scaled-down secret images.

[0051] The carrier image is divided into N non-overlapping blocks, and the reversible image hiding module sequentially hides the reduced secret image into the corresponding non-overlapping blocks to obtain N hidden image blocks; wherein, the size of the reduced secret image is the same as the size of the non-overlapping block;

[0052] The N dense image blocks are sequentially merged to obtain the predicted dense image;

[0053] The image steganalysis network model is trained based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model.

[0054] Image steganography is performed using the trained image steganography network model to obtain a cryptic output image.

[0055] The aforementioned large-scale image steganography method, apparatus, computer device, and storage medium based on reversible neural networks downsample the secret image to a corresponding factor through a reversible image scaling module; simultaneously, the carrier image is segmented into non-overlapping blocks of the same size as the downsampled secret image; the generated downsampled secret image and the carrier image are input into a reversible image hiding module to generate a steganographic image. This invention, by introducing a reversible image scaling transformation and fully utilizing the bidirectional nature of reversible neural networks, overcomes problems such as mutual interference of information and color between secret images, significantly improving the quantity of image steganography and the quality of decrypted image recovery. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a large-scale image steganography method based on a reversible neural network in one embodiment.

[0057] Figure 2 This is a schematic diagram of a large-scale image steganography model based on a reversible neural network in one embodiment, wherein, Figure 2 (a) represents the downsampling and hiding process. Figure 2 (b) The process of reconstructing and upsampling the decrypted image;

[0058] Figure 3 This is a schematic diagram illustrating the results of a method for comparing two hidden images in one embodiment.

[0059] Figure 4 This is a comparison of the results of steganography of two secret images in one embodiment, wherein, Figure 4 (a) is a graph showing the results of the GT method. Figure 4 (b) is a graph showing the results of the Weng method. Figure 4 (c) is a graph showing the results of the Baluja method. Figure 4 (d) is a graph showing the results of the ISN method. Figure 4 (e) is a graph showing the results of the DeepMIH method. Figure 4 (f) is a result diagram of the InvMIHNet method of the present invention;

[0060] Figure 5 A visualization of the result of steganography of 6 secret images in one embodiment;

[0061] Figure 6 A visualization of the result of steganography of 8 secret images in one embodiment;

[0062] Figure 7 A visualization of the result of steganography of 9 secret images in one embodiment;

[0063] Figure 8 A visualization of the result of steganography of 16 secret images in one embodiment;

[0064] Figure 9 This is a structural block diagram of a large-scale image steganography device based on a reversible neural network in one embodiment;

[0065] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0067] In one embodiment, such as Figure 1 As shown, a large-scale image steganography method based on a reversible neural network is provided, including the following steps:

[0068] Step 102: Obtain the carrier image and N secret images to be hidden, and input the carrier image and secret images into the pre-designed image steganography network model.

[0069] The image steganography network model includes a reversible image scaling module and a reversible image hiding module.

[0070] The image steganography network model proposed in this invention is summarized as follows: Figure 2 As shown, taking the hiding and restoring of four secret images as an example, the model consists of a reversible image scaling module and a reversible image hiding module.

[0071] Step 104: The secret image is downsampled using the reversible image scaling module to obtain N scaled-down secret images.

[0072] First, the reversible image scaling module is used to process N secret images. By downsampling, N smaller secret images are obtained. and variables unrelated to the image The basic information of the secret image is compressed into the downsampled secret image. middle.

[0073] Step 106: Divide the carrier image into N non-overlapping blocks, and use the reversible image hiding module to hide the reduced secret image into the corresponding non-overlapping blocks in sequence to obtain N secret image blocks.

[0074] The size of the scaled-down secret image is the same as the size of the non-overlapping block.

[0075] After the carrier image is segmented into N non-overlapping blocks, the k-th scaled-down secret image is hidden using a reversible image hiding module. Hide to the kth block In the middle, and generate the k-th dense image patch. And a variable unrelated to the image. By merging the carrier image blocks in the corresponding order, a hidden image x containing N secret images can be obtained. stgeo .

[0076] Step 108: Merge the N dense image blocks sequentially to obtain the predicted dense image.

[0077] Step 110: Train the image steganalysis network model based on the predicted dense image and the preset loss function to obtain the trained image steganalysis network model.

[0078] The training objective of this invention includes two aspects: (i) minimizing the difference between the reconstructed decrypted image and the original secret image; (ii) minimizing the difference between the carrier image and the decrypted image. Therefore, the loss function consists of two parts: Used to guide the learning process of secret images, and To improve the quality of dense images, the overall loss function can be expressed as:

[0079]

[0080] Specifically This loss function is used to guide the downsampled secret image to contain sufficient information, while simultaneously ensuring the quality of the decrypted image after upsampling. It comprises three parts: (i) a recovery loss function, used to improve the quality of the upsampled decrypted image; (ii) a downsampling loss function, ensuring that the downsampled secret image is as similar as possible to the reference image obtained through bicubic interpolation; and (iii) a reconstruction loss function, used to ensure that the decrypted image reconstructed by the image hiding module is as consistent as possible with the input secret image. Furthermore, a distribution matching loss function is set, using JS divergence to limit the distributions of the secret image and the decrypted image to overlap as much as possible. Therefore, This can be further expanded to represent:

[0081]

[0082] Where p(·) represents the data distribution, and λ1, λ2 and λ3 represent the weight ratios of the recovery loss function, the downsampling loss function and the reconstruction loss function, respectively.

[0083] Steganography loss function Used to ensure dense images and original carrier image The differences between them are difficult to distinguish, ensuring the stealth of the steganography. At the same time, a loss function is set to ensure similarity in low-frequency regions, allowing the secret information to be hidden as much as possible in higher-frequency regions that are harder to perceive.

[0084]

[0085] The subscript ll indicates the low-frequency information obtained through Haar wavelet transform.

[0086] High-quality image steganography relies on efficient image scaling. The proposed image steganography network model involves the sequential application of the forward and inverse processes of a reversible image scaling module and a reversible image hiding module, which is difficult to optimize simultaneously. Therefore, instead of training the entire network from scratch, a two-stage training approach is adopted. In the first stage, the reversible image scaling module and the reversible image hiding module are trained separately for 30K iterations. In the second stage, the reversible image scaling module and the reversible image hiding module are jointly trained for another 20K iterations.

[0087] Step 112: Steganography is performed using the trained image steganography network model to obtain the output cryptic image.

[0088] In the aforementioned large-scale image steganography method based on reversible neural networks, the secret image is downsampled to a corresponding factor using a reversible image scaling module; simultaneously, the carrier image is segmented into non-overlapping blocks of the same size as the downsampled secret image; the generated downsampled secret image and the carrier image are input into a reversible image hiding module to generate a steganographic image. This invention employs a divide-and-conquer strategy, hiding the scaled-down secret image within non-overlapping blocks, thereby significantly improving the hiding capability of the carrier image and the recovery quality of the decrypted image. By introducing a reversible image scaling transformation and fully utilizing the bidirectional nature of the reversible neural network, it overcomes problems such as mutual interference of information and color between secret images, greatly increasing the number of steganographic representations and the quality of the decrypted image recovery.

[0089] In one embodiment, the method further includes: acquiring a secret image to be decrypted; decrypting the secret image to be decrypted through the reverse process of the reversible image hiding module to obtain N initial decrypted images; upsampling the N initial decrypted images through the reversible image scaling module to obtain N decrypted images; the image resolution of the decrypted images is the same as that of the original secret image.

[0090] Based on the bidirectional nature of reversible neural networks, the process of decrypting and reconstructing the image is the inverse operation of the reversible image hiding module and the reversible image scaling module. That is, the decrypted image is generated through the inverse process of the reversible image scaling module, and then the decrypted image is upsampled through the inverse process of the reversible image hiding module, finally reconstructing a decrypted image with the same resolution as the original secret image.

[0091] In one embodiment, the forward process of the reversible image scaling module includes a decomposition block and multiple reversible blocks. The secret image is downsampled by the convolutional layer in the decomposition block of the reversible image scaling module to obtain low-frequency information and high-frequency information generated by the decomposition block. The reversible blocks in the reversible image scaling module exchange information based on the low-frequency information and high-frequency information generated by the decomposition block to obtain N scaled-down secret images and a variable unrelated to an image.

[0092] The reverse process of the reversible image scaling module includes multiple reversible blocks and a synthesis block. The reversible blocks in the reversible image scaling module exchange information between the initial decrypted image and a randomly sampled variable to obtain low-frequency and high-frequency information output by the reversible block's reverse process. The synthesis block in the reversible image scaling module upsamples the low-frequency and high-frequency information output by the reversible block's reverse process to obtain N restored decrypted images.

[0093] The forward process of the reversible image hiding module includes a discrete wavelet transform block and multiple reversible blocks. The discrete wavelet transform block in the reversible image hiding module decomposes the reduced secret image and non-overlapping blocks into low-frequency information and high-frequency information. The reversible blocks in the reversible image hiding module exchange information based on the low-frequency information and high-frequency information obtained from the discrete wavelet transform block decomposition, generating N dense image blocks and a variable independent of the image.

[0094] The reverse process of the reversible image hiding module includes multiple reversible blocks and an inverse discrete wavelet transform block. The reversible blocks in the reversible image hiding module exchange information between the encrypted image to be decrypted and a randomly sampled variable to obtain the low-frequency and high-frequency information output by the reversible blocks. The low-frequency and high-frequency information output by the reversible blocks are then reconstructed by the inverse discrete wavelet transform block in the reversible image hiding module to obtain N reconstructed initial decryption images.

[0095] Specifically, the forward and reverse processes of the reversible image scaling module are represented as follows: and Where η represents the learnable parameters in the reversible image scaling module. It is sampled from a normal distribution. Similarly, the forward and inverse processes of the reversible image hiding module can be represented as follows: and Where θ represents the learnable parameters in the reversible image hiding module. Sampling is performed from a normal distribution. For simplicity, irrelevant parts are omitted, and the hiding process is simplified as follows: and

[0096] In summary, the learning objective of the image steganalysis network model proposed in this invention is expressed by formula (1):

[0097]

[0098] Where θ and η represent the learnable parameters in the reversible image hiding module and the reversible image scaling module, respectively; λ is set to balance the effects of hiding and restoring the secret image; x c and x s These represent the carrier image and the secret image, respectively. and They are all sampled from a predefined Gaussian distribution and are used to supplement the information lost during the steganography process.

[0099] Invertible blocks are fundamental components of reversible image hiding modules and reversible image scaling modules. Because the forward and backward processes of invertible blocks share parameters, reversible neural networks can be used to learn nonlinear transformations and can reconstruct the function of the inverse process without loss. If we consider… and Defined as the output of the i-th reversible block, and simultaneously set in Since the decomposition blocks are based on the downsampling task transformation, the forward process of the i-th invertible block can be represented as:

[0100]

[0101]

[0102] in, and These represent the low-frequency and high-frequency information generated by the i-th reversible block, respectively. ρ(·) and ψ(·) both represent dense networks. By utilizing the bidirectional nature of reversible blocks, the reverse process of the i-th reversible block can be expressed as:

[0103]

[0104]

[0105] The goal of the reversible image scaling module is to scale the reversible secret image. Assuming the carrier image and the secret image are the same size, when m×n secret images need to be hidden, the secret images need to be downsampled by the reversible image scaling module, where the downsampling factor is m×n. During the downsampling process, the downsampled secret image generated by the forward process of the reversible image scaling module must contain as much secret information as possible, and it must be ensured that the reconstructed decrypted image is as similar as possible to the original secret image during the reverse process.

[0106] The forward process of the reversible image scaling module includes a decomposition block. as well as There are reversible blocks, and the reverse process includes... One reversible block and one composite block A secret image is obtained by breaking down blocks. Convolutional layers in Will Decomposed into It includes the high-frequency domain and low frequency domain The output image can be synthesized using blocks. The deconvolutional layer in the image transforms information from the high-frequency and low-frequency domains back to the original image size. Finally, the output is passed through a quantization block, which ensures that the pixel values ​​of the image are integers within the range [0, 255].

[0107] The forward hiding process of the reversible image hiding module involves downsampling m×n secret images to a multiple of m×n of the original image and then hiding them within m×n non-overlapping carrier image blocks segmented into the carrier image. The reverse process reconstructs the decrypted image from the hidden image. Specifically, the reversible image hiding module contains a Discrete Wavelet Transform (DWT) / Inverse Discrete Wavelet Transform (IDWT) block and G reversible blocks. DWT decomposes the input secret image into high-frequency and low-frequency information domains, then hides the secret information within the carrier image. In this paper, the Haar Wavelet Transform is used by default. During the forward propagation of the reversible image hiding module, the downsampled secret image... and carrier image blocks First, the image is decomposed into information of different domains using the DWT (Decomposition Theory Wrapper). This information is then input into G reversible blocks for information exchange, generating a corresponding cryptic block and a variable independent of the image. The cryptic image is then processed by a quantization block, ensuring that the pixel values ​​are integers within the range [0, 255]. During the reverse propagation of the reversible image hiding module, the cryptic image block and a randomly sampled variable are used as inputs through the G reversible blocks and the IDWT to obtain the reconstructed carrier image. and the decrypted image The output decrypted image will then be upsampled through the reverse propagation process of the reversible image scaling module to obtain the final decrypted image.

[0108] In one specific embodiment, an experiment was conducted on a large-scale image steganography method based on a reversible neural network, including:

[0109] Experimental setup:

[0110] (1) Dataset selection

[0111] In this embodiment, the training dataset chosen is the DIV2K training dataset, which contains 800 training images with a resolution of 1024×1024. The test images include 100 DIV2K validation images, plus 1,000 randomly selected 256×256 images from the ImageNet validation set and 5,000 256×256 images from the COCO test set.

[0112] (2) Comparison Method

[0113] Five comparison methods were selected: Weng, Baluja, and UDH. Weng, Baluja, and UDH are all based on deep convolutional neural networks and can hide and recover two secret images. DeepMIH and ISN are methods for steganography of large-capacity images based on invertible neural networks and can hide and recover four secret images. For a fair comparison, all methods were trained and tested on the same dataset.

[0114] (3) Implementation details

[0115] The optimizer used in this embodiment is Adam, with an initial learning rate of 2×10⁻⁶. -4 The learning rate is halved every 10,000 iterations. The size of the training image patch is set to 144×144. The weight coefficients are set to λ1=1, λ2=4, and λ3=5 respectively.

[0116] (4) Measurement indicators

[0117] The same metrics used in related work were employed, including: Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM), Root-Mean-Square Error (RMSE), and Mean Absolute Error (MAE).

[0118] Experimental results:

[0119] (1) Steganographically write two secret images

[0120] When hiding two secret images, a reversible image scaling module is used to reduce the secret image by a factor of 2×1. This method achieves an initial PSNR of 47.79 dB on the hidden image and the carrier image, and an initial PSNR of 35.47 dB on the secret image and the decrypted image. In comparison, ISN achieves a PSNR of 45.95 dB on the hidden image and carrier image pair, and 36.88 dB on the secret image and decrypted image pair, while DeepMIH achieves a PSNR of 40.62 dB on the hidden image and carrier image pair, and 41.70 dB on the secret image and decrypted image pair.

[0121] As can be seen, different methods emphasize different aspects of both hiding and restoring the secret image. Since the hidden image contains information from both the secret image and the carrier image, there must be a trade-off between the quality of the hidden image / carrier image pair and the quality of the secret image / decrypted image pair. Therefore, it can be inferred that given an image steganography method, the quality of the decrypted image cannot be improved without sacrificing the hiding effect of the secret image. The results were fine-tuned by slightly increasing the regularization parameter λ1 and halving the learning rate. Figure 3 As can be seen, this method can fit the linear relationship very well.

[0122] Here, every point on the curve is considered equivalent, and these points form the Pareto boundary. Therefore, in the following sections, all methods are evaluated based on the average score of the hidden and carrier images versus the secret and decrypted images. This new metric better evaluates the overall performance of different image steganography methods in terms of hiding and recovery.

[0123] Table 1. Evaluation results of hiding two secret images on different datasets

[0124]

[0125] Table 1 shows the results of hiding two secret images evaluated on DIV2K, COCO, and ImageNet. Evaluation metrics include mean PSNR, SSIM, RMSE, and MAE. In addition to the DIV2K dataset, the performance of our method was also tested on the COCO and ImageNet datasets to evaluate its transferability across datasets. As can be seen from the table, our method achieves state-of-the-art performance on both DIV2K and ImageNet datasets, and second-best performance on COCO. Figure 4 This is a comparison of the results of steganography on two secret images. Figure 4 (a) is a graph showing the results of the GT method. Figure 4 (b) is a graph showing the results of the Weng method. Figure 4 (c) is a graph showing the results of the Baluja method. Figure 4 (d) is a graph showing the results of the ISN method. Figure 4 (e) is a graph showing the results of the DeepMIH method. Figure 4 (f) is a result diagram of the InvMIHNet method of the present invention. Figure 4 From top to bottom, they are: the secret image, the difference between the carrier image and the secret image, the first decrypted image reconstructed, the difference between the secret image and the reconstructed image, the second decrypted image reconstructed, and the difference between the carrier image and the original secret image.

[0126] (2) Steganographically write 4 secret images

[0127] Table 2 lists the results of hiding four secret images. Compared with ISN and DeepMIH, the method proposed in this invention achieves the best image steganography results. Specifically, compared with the steganography results of ISN, this method improves PSNR by 5.6dB, 1.2dB, and 5.1dB on the DIV2K, COCO, and ImageNet datasets, respectively.

[0128] Table 2. Evaluation results of hiding four secret images on different datasets

[0129]

[0130] It is worth noting that although DeepMIH achieves the second-best performance, it requires approximately four times more learnable parameters than our proposed method. The specific parameter counts for each method are shown in Table 3. Furthermore, it should be noted that in DeepMIH, each secret image corresponds to a separate invertible neural network, so the number of parameters increases linearly with the number of secret images, making it difficult to apply to image steganography with larger capacities. In contrast, the image steganography method proposed in this invention can successfully hide and recover a larger number of secret images with a slightly increased number of parameters, thus significantly reducing training time compared to DeepMIH.

[0131] Table 3. Number of parameters required for training for three steganography methods (* indicates estimated number of parameters)

[0132]

[0133] (3) Steganographically write N (N>4) secret images

[0134] Next, steganography experiments were conducted on more secret images, namely N = 6, 8, 9, and 16, which is a number of steganography results never achieved in related works. Since no other comparative method can achieve the same number of images as in this experiment, only Table 4 shows the image steganography results of this method.

[0135] Table 4. Results of Large-Capacity Image Steganography

[0136]

[0137] As can be seen, the effectiveness of image hiding and reconstruction gradually decreases with the increase in the number of secret images. Even when the number of secret images increases to 16, this method can still achieve a decryption image reconstruction effect of approximately 30.6dB PSNR, indicating that this method provides a strong baseline for large-capacity image steganography.

[0138] Figures 5-8 The steganography results for 6, 8, 9, and 16 secret images were visualized. Within the solid lines, from top to bottom, are the carrier image, the steganographic image, and the difference between the two. Within the dashed lines, the steganography results for the secret images are shown, from top to bottom: the secret image to be hidden, the reconstructed decrypted image, and the difference between the two. The differences between the decrypted and secret images reveal that, due to the loss of high-frequency information during image scaling and steganography, most of these differences pertain to high-frequency domain information.

[0139] Experimental results show that the proposed model outperforms state-of-the-art methods in the following aspects: imperceptibility of dense images, recovery accuracy of decrypted images, and security of concealment, further demonstrating the outstanding ability of image steganography network models in ultra-large capacity image steganography.

[0140] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0141] In one embodiment, such as Figure 9 As shown, a large-scale image steganography device based on a reversible neural network is provided, comprising: an image input module 902, a downsampling module 904, an image hiding module 906, a model prediction module 908, a model training module 910, and a model application module 912, wherein:

[0142] The image input module 902 is used to acquire a carrier image and N secret images to be hidden, and input the carrier image and secret images into a pre-designed image steganography network model; the image steganography network model includes a reversible image scaling module and a reversible image hiding module;

[0143] The downsampling module 904 is used to downsample the secret image through the reversible image scaling module to obtain N scaled-down secret images;

[0144] The image hiding module 906 is used to divide the carrier image into N non-overlapping blocks, and to hide the reduced secret image into the corresponding non-overlapping blocks in sequence through the reversible image hiding module, thereby obtaining N hidden image blocks; wherein, the size of the reduced secret image is the same as the size of the non-overlapping block.

[0145] The model prediction module 908 is used to sequentially merge N dense image blocks to obtain the predicted dense image;

[0146] The model training module 910 is used to train the image steganalysis network model based on the predicted dense image and the preset loss function to obtain the trained image steganalysis network model.

[0147] The model application module 912 is used to perform image steganography using a trained image steganography network model to obtain a steganographic output image.

[0148] The model application module 912 is also used to acquire the secret image to be decrypted; the secret image to be decrypted is decrypted through the reverse process of the reversible image hiding module to obtain N initial decrypted images; the N initial decrypted images are upsampled through the reversible image scaling module to obtain N decrypted images; the image resolution of the decrypted images is the same as that of the original secret image.

[0149] The downsampling module 904 is also used to downsample the secret image through the convolutional layer in the decomposition block of the reversible image scaling module to obtain low-frequency and high-frequency information generated by the decomposition block; and to exchange information based on the low-frequency and high-frequency information generated by the low-frequency and high-frequency information generated by the low-frequency and high-frequency information generated by the decomposition block through the reversible block in the reversible image scaling module to obtain N scaled-down secret images and variables independent of one image.

[0150] The model application module 912 is also used to exchange information between the initial decrypted image and a randomly sampled variable through the reversible block in the reversible image scaling module, to obtain the low-frequency and high-frequency information output by the reversible block in reverse process; and to obtain N restored decrypted images by upsampling the low-frequency and high-frequency information output by the synthesis block in the reversible image scaling module based on the low-frequency and high-frequency information output by the reversible block in reverse process.

[0151] The image hiding module 906 is also used to decompose the reduced secret image and non-overlapping block into low-frequency information and high-frequency information through the discrete wavelet transform block in the reversible image hiding module; and to exchange information based on the low-frequency information and high-frequency information obtained by the discrete wavelet transform block through the reversible block in the reversible image hiding module, thereby generating N dense image blocks and a variable unrelated to the image.

[0152] The model application module 912 is also used to exchange information between the encrypted image to be decrypted and a randomly sampled variable through the reversible block in the reversible image hiding module, so as to obtain the low-frequency information and high-frequency information output by the reversible block; and to reconstruct the low-frequency information and high-frequency information output by the reversible block through the inverse discrete wavelet transform block in the reversible image hiding module, so as to obtain N reconstructed initial decryption images.

[0153] The model training module 910 is also used to iteratively train the reversible image scaling module and the reversible image hiding module respectively based on the predicted dense image and the preset loss function, and then jointly train the reversible image scaling module and the reversible image hiding module iteratively; wherein, the preset loss function is:

[0154]

[0155] in, The secret image used to guide downsampling contains sufficient information, including the recovery loss function, the downsampling loss function, and the reconstruction loss function; This is the steganalysis loss function, used to constrain the difference between the output cryptic image and the original carrier image to be as small as possible.

[0156] Specific limitations regarding large-scale image steganography devices based on reversible neural networks can be found in the limitations of large-scale image steganography methods based on reversible neural networks described above, and will not be repeated here. Each module in the aforementioned large-scale image steganography device based on reversible neural networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0157] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a large-scale image steganography method based on a reversible neural network. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0158] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0159] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiment.

[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0161] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0162] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0163] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A large-scale image steganography method based on reversible neural network, characterized in that, The method includes: acquiring a carrier image and a secret image to be hidden, inputting the carrier image and the secret image into a pre-designed image steganography network model; the image steganography network model comprises a reversible image scaling module and a reversible image hiding module; down-sampling the secret image by the reversible image scaling module to obtain a zoomed-in secret image; dividing the carrier image into non-overlapping blocks, hiding the reduced secret image into the corresponding non-overlapping blocks in sequence by the reversible image hiding module to obtain a plurality of stego image blocks; wherein the size of the reduced secret image is consistent with the size of the non-overlapping blocks; The The Zhang image blocks are sequentially merged to obtain a predicted Zhang image. The image steganalysis network model is trained based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model. Image steganography is performed using the trained image steganography network model to obtain a densely packed output image; The method further includes: Obtain the encrypted image to be decrypted; decrypting the image containing secret to be decrypted by inverse process of the reversible image hiding module, obtaining Zhang Chujie decrypted image; The reversible image zooming module is used to zoom the decrypted image to obtain a decrypted image with the same image resolution as the original secret image. The decrypted image is up-sampled to obtain a decrypted image with the same image resolution as the original secret image. The decrypted image is up-sampled to obtain a decrypted image with the same image resolution as the original secret image. The forward process of the reversible image hiding module includes a discrete wavelet transform block and multiple reversible blocks, and the reduced secret image is sequentially hidden into corresponding non-overlapping blocks through the reversible image hiding module to obtain a secret-containing image block. The reduced secret image and the non-overlapping block are decomposed into low-frequency information and high-frequency information by the discrete wavelet transform block in the reversible image hiding module; The low frequency information and the high frequency information obtained by the reversible block in the reversible image hiding module according to the discrete wavelet transform block decomposition are exchanged for information, and one image block containing a secret image and one variable irrelevant to the image are generated ​ The inverse process of the reversible image hiding module includes a plurality of reversible blocks and an inverse discrete wavelet transform block. The encrypted image to be decrypted is decrypted by the inverse process of the reversible image hiding module to obtain Zhang Chujie decrypts the image, comprising: The reversible block in the reversible image hiding module exchanges information between the encrypted image to be decrypted and a randomly sampled variable to obtain the low-frequency and high-frequency information output by the reversible block. The low frequency information and the high frequency information outputted by the reversible block are reconstructed by the inverse discrete wavelet transform block in the reversible image hiding module, and the reconstructed Zhang Chujie decrypts the image.

2. The method of claim 1, wherein, The forward process of the reversible image scaling module includes a decomposition block and a plurality of reversible blocks, and the secret image is down-sampled by the reversible image scaling module to obtain a secret image after zooming out, comprising: The secret image is downsampled by the convolutional layer in the decomposition block of the reversible image scaling module to obtain the low-frequency and high-frequency information generated by the decomposition block. The low frequency information and the high frequency information generated by the reversible block in the reversible image zooming module are exchanged to obtain The secret image after zooming is irrelevant to a variable of an image.

3. The method of claim 2, wherein, The inverse process of the reversible image scaling module includes a plurality of reversible blocks and a synthesis block, and the image is scaled by the reversible image scaling module up-sampling the decrypted image to obtain a decrypted image, comprising: The reversible block in the reversible image scaling module exchanges information between the initially decrypted image and a randomly sampled variable to obtain low-frequency and high-frequency information output by the reverse process of the reversible block. The low frequency information and the high frequency information outputted by the inverse process of the reversible block in the synthesis block of the reversible image zooming module are up-sampled to obtain The decrypted image is reduced by Zhang.

4. The method according to any one of claims 1 to 3, characterized in that, In the reversible image zoom module and the reversible image hiding module, the forward process of the first reversible block is represented as: ​ No. The reverse process of a reversible block is represented as follows: wherein, and respectively represent low frequency information and high frequency information of the th reversible block generation, , and all represent dense networks.

5. The method of claim 1, wherein, The image steganalysis network model is trained based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model, including: Based on the predicted dense image and a preset loss function, the reversible image scaling module and the reversible image hiding module are iteratively trained separately, and then jointly trained iteratively; wherein, the preset loss function is: wherein, The guide for down-sampling the secret image contains sufficient information, including a recovery loss function, a down-sampling loss function, and a reconstruction loss function; The steganography loss function is used to constrain the difference between the output stego image and the original carrier image to be as small as possible.

6. A large-scale image steganography device based on reversible neural networks, characterized by, The apparatus comprising the method according to any one of claims 1 to 5, wherein the apparatus includes: an image input module, configured to acquire a carrier image and a secret image to be hidden, input the carrier image and the secret image into a pre-designed image steganography network model; the image steganography network model comprises a reversible image scaling module and a reversible image hiding module; a down-sampling module, configured to down-sample the secret image by the reversible image scaling module to obtain a zoomed-in secret image; an image hiding module, configured to split the carrier image into non-overlapping blocks, hide the reduced secret image into the corresponding non-overlapping blocks in sequence by the reversible image hiding module, to obtain a secret-containing image block; wherein the size of the reduced secret image is consistent with the size of the non-overlapping block. a model prediction module configured to predict the image based on the model the image block sequence is merged to obtain a predicted image with the secret The model training module is used to train the image steganalysis network model based on the predicted dense image and a preset loss function to obtain a trained image steganalysis network model. The model application module is used to perform image steganography using the trained image steganography network model to obtain the output steganographic image. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

Images