Image steganography method and system based on frequency domain transformation and reversible neural network
By combining frequency domain transformation and reversible neural networks, the shortcomings of insufficient utilization of frequency domain features and reversibility design in existing technologies are solved, achieving lossless recovery of secret information and high concealment, thus improving the concealment and recovery accuracy of steganographic images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-17
AI Technical Summary
Existing image steganography methods do not make sufficient use of frequency domain features, resulting in an imbalance between concealment and recovery accuracy. Furthermore, the lack of strict constraints on reversibility design makes steganographic images easy to detect and distorts secret information.
A method based on frequency domain transformation and reversible neural networks is adopted. The secret image and the carrier image are decomposed into multiple sub-bands through wavelet transform. The reversible network of the Glow architecture is used for feature fusion and reconstruction. Combined with a multi-loss constraint mechanism, the reversibility and concealment of the steganography process are ensured.
It achieves lossless recovery of secret information and high concealment of steganalytical images, improves the visual consistency between carrier images and steganalytical images, and reduces the detection risk of steganalysis.
Smart Images
Figure CN121883263A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of information hiding and digital image processing, and specifically relates to an image steganography method and system based on frequency domain transformation and reversible neural networks. Background Technology
[0002] The core of image steganography is to embed secret information into a carrier image without being detected, and to accurately extract the secret information from the steganographic image. Traditional methods have two major limitations: spatial domain methods (such as least significant bit substitution) directly modify pixel values, which can easily produce visual traces in smooth areas, resulting in poor concealment; single transform domain methods (such as relying solely on DCT or wavelet transform) depend on manually selected embedding positions, which cannot adapt to the texture features of the carrier image, leading to significant differences in the frequency domain statistical characteristics between the steganographic image and the carrier image, making it easy to be detected by steganalysis tools.
[0003] In recent years, deep learning-based steganography methods have improved the automation of embedding through end-to-end learning, but key shortcomings still exist: First, the feature fusion mechanism is singular, mostly performing feature splicing in the pixel domain or a single transform domain, failing to fully utilize the hierarchical characteristics of the frequency domain (low-frequency regions are visually sensitive, and high-frequency regions have high redundancy), resulting in unreasonable embedding positions of secret information and poor visual consistency of steganographic images; Second, the reversibility design is insufficient, with most models adopting an encoder-decoder architecture, which easily loses feature information during forward embedding, leading to distortion of secret information during reverse extraction, failing to meet the requirement of strict reversibility of "embedding-extraction".
[0004] The bottlenecks of the existing technology are specifically manifested as follows: (1) Insufficient utilization of frequency domain features, and difficulty in balancing "low-frequency concealment" and "high-frequency embedding capacity" in a single transform domain. Modifications in the low-frequency region are easily detected, while excessive embedding in the high-frequency region reduces the accuracy of secret recovery; (2) Defects in the design of reversible network architecture. Traditional cascaded networks lack strict reversibility constraints (such as channel dimension consistency and transform function reversibility), resulting in information loss; (3) One-sided design of loss function, focusing only on pixel-level errors and ignoring the statistical consistency of video domain sub-bands (especially low-frequency sub-bands), resulting in steganalysis images being easily detected in the frequency domain.
[0005] To address the aforementioned issues, existing technologies have not yet formed a solution for "deep integration of frequency domain hierarchical embedding and reversible networks": poor compatibility between frequency domain transformation and reversible networks, and mismatch in feature dimensions can easily destroy reversibility; reversible networks do not combine frequency domain characteristics to guide the embedding of secret information into redundant regions of the carrier, resulting in an imbalance between concealment and recovery accuracy. Summary of the Invention
[0006] This invention addresses the technical shortcomings of existing image steganography methods, such as insufficient utilization of frequency domain features and an imbalance between reversibility and concealment. It proposes an image steganography method based on frequency domain transformation and a reversible neural network. The method involves writing a secret image into a carrier image at the transmitting end and sending it to the receiving end, where the secret image is reconstructed from the received image. Specifically, it includes the following steps:
[0007] At the transmitting end, each channel of the 3-channel secret image and carrier image is decomposed into one low-frequency sub-band and three high-frequency sub-bands through wavelet transform;
[0008] First, all subbands of the secret image and the carrier image are spliced together to form 12-channel fusion features. Then, the 12-channel fusion features of the two images are spliced together to form 24-channel fusion features.
[0009] The 24-channel fused features are input into a reversible network for forward processing to obtain 12-channel stegographic carrier frequency domain features.
[0010] The frequency domain features of the 12-channel stegographic carrier are used to generate a pixel-domain stegographic image through inverse wavelet transform;
[0011] At the receiving end, the pixel-domain steganalysis image is decomposed into a low-frequency sub-band and three high-frequency sub-bands by wavelet transform. Each sub-band is superimposed with 12-channel noise-filled features based on Gaussian distribution to form 24-channel input features.
[0012] The reversible network performs inverse processing on the 24-channel input features, splitting the 24-channel input features into 12-channel secret frequency domain features;
[0013] The 12-channel secret frequency domain features are reconstructed into a secret restored image through inverse wavelet transform.
[0014] This invention also proposes an image steganography system based on frequency domain transformation and reversible neural networks, used to implement an image steganography method based on frequency domain transformation and reversible neural networks, comprising:
[0015] The wavelet transform module is used to perform frequency domain decomposition of the carrier image, secret image, and stegana.
[0016] The inverse wavelet transform module is used to reconstruct the image from frequency domain features to pixel domain;
[0017] The reversible network consists of 16 GlowBlocks connected in series. Each GlowBlock includes a cascaded activation normalization layer, a reversible 1×1 convolutional layer, and an affine coupling layer, which are used to achieve reversible transformation of 24-channel fused features.
[0018] Compared with existing technologies, the core innovations of this invention are as follows:
[0019] 1. Strictly reversible design based on Glow architecture: Through the collaborative design of ActNorm, reversible 1×1 convolution and coupled layer subnets, the strict reversibility of 16 GlowBlock cascaded networks is guaranteed, and the secret information can be recovered without distortion.
[0020] 2. Frequency Domain Layered Embedding of Discrete Wavelet Transform: Using DWT to decompose the high and low frequency sub-bands of the carrier and the secret image, the secret information is guided to be adaptively embedded into the high frequency texture region and low frequency redundant region of the carrier, thereby improving the stealth of the steganography.
[0021] 3. Multi-loss collaborative constraint mechanism: It integrates pixel domain and frequency domain losses, and simultaneously optimizes the visual consistency between the stegana and the carrier and the secret recovery accuracy, balancing the contradiction between "concealment and reversibility". Attached Figure Description
[0022] Figure 1 This is a diagram illustrating the overall framework of an image steganography method based on frequency domain transformation and reversible neural networks according to the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] This invention proposes an image steganography method based on frequency domain transformation and reversible neural networks. The method involves writing a secret image into a carrier image at the transmitting end and sending it to the receiving end, where the secret image is reconstructed from the received image. Specifically, the method includes the following steps:
[0025] At the transmitting end, each channel of the 3-channel secret image and carrier image is decomposed into one low-frequency sub-band and three high-frequency sub-bands through wavelet transform;
[0026] First, all subbands of the secret image and the carrier image are spliced together to form 12-channel fusion features. Then, the 12-channel fusion features of the two images are spliced together to form 24-channel fusion features.
[0027] The 24-channel fused features are input into a reversible network for forward processing to obtain 12-channel stegographic carrier frequency domain features.
[0028] The frequency domain features of the 12-channel stegographic carrier are used to generate a pixel-domain stegographic image through inverse wavelet transform;
[0029] At the receiving end, the pixel-domain steganalysis image is decomposed into a low-frequency sub-band and three high-frequency sub-bands by wavelet transform. Each sub-band is superimposed with 12-channel noise-filled features based on Gaussian distribution to form 24-channel input features.
[0030] The reversible network performs inverse processing on the 24-channel input features, splitting the 24-channel input features into 12-channel secret frequency domain features;
[0031] The 12-channel secret frequency domain features are reconstructed into a secret restored image through inverse wavelet transform.
[0032] like Figure 1 In this invention, a carrier image with 3 channels and a size of H×W and a secret image with 3 channels and a size of H×W are processed by wavelet transform to obtain one high-frequency sub-feature map and three low-frequency sub-feature maps. Each sub-feature map has 3 channels and a size of H / 2×W / 2. All these feature maps are spliced along the channels to form a 12-channel feature map. The 12-channel feature maps of the carrier image and the secret image are spliced along the channel dimension to form a 24-channel feature map, which serves as the input to the forward process of the reversible neural network. The reversible neural network consists of 16 cascaded GlowBlocks. In the forward process, the reversible neural network outputs image features with 24 channels and a size of H / 2×W / 2. From these image features, [the following text is incomplete and requires further context:] The steganalysis features with 12 channels and a size of H / 2×W / 2 are obtained. These features are then subjected to inverse wavelet transform to obtain the steganalysis image. A high-frequency sub-feature map and three low-frequency sub-feature maps are extracted from the steganalysis image through wavelet transform. These sub-feature maps are then concatenated to obtain a feature map with 12 channels and a size of H / 2×W / 2. This feature map is then concatenated with a feature map with 12 channels and a size of H / 2×W / 2 generated by noise in the channel dimension. This concatenation is then used as the input for the inverse process of the reversible neural network. The inverse process of the reversible neural network outputs an image feature with 24 channels and a size of H / 2×W / 2. This feature is then split to obtain a feature with 12 channels and a size of H / 2×W / 2. This feature is then subjected to inverse wavelet transform to obtain the secret image.
[0033] This embodiment is implemented based on the PyTorch framework, with an NVIDIA GPU P100 (16GB of video memory) as the hardware environment and Python 3.11, PyTorch 2.6.0, and CUDA 12.4 as the software environment.
[0034] 1. Dataset Preparation
[0035] The DIV2K dataset (containing 800 training images and 100 validation images) was used. During training, the image size was uniformly adjusted to 256×256×3, and the data was enhanced by random horizontal flipping, vertical flipping, and random cropping. The input images were normalized to the range [0,1] by ToTensor transformation.
[0036] 2. Network parameter settings
[0037] Frequency domain transformation: The wavelet transform uses the Haar kernel, with a level 1 decomposition and a subband size of 128×128×3;
[0038] Reversible network: 16 GlowBlocks concatenated, ActNorm parameters automatically initialized, reversible 1×1 convolution weights are orthogonal matrices, and the last convolution layer of GlowCouplingSubnet is initialized to 0;
[0039] Training parameters: batch_size=12, initial learning rate 3e-5, optimizer Adam (betas=(0.5,0.999), weight_decay=1e-5), gradient clipping threshold 1.0, and learning rate dynamically adjusted through a periodic decay scheduling strategy.
[0040] 3. Forward Steganography and Reverse Extraction Process
[0041] Forward process: The cover image and the secret image are decomposed into 12-channel frequency domain features by DWT and concatenated into a 24-channel input; after the input is processed by StegoNet, it is split to obtain 12-channel stegographic carrier features, which are then reconstructed into a stegographic image by IWT.
[0042] Reverse process: The steganalog image (stego) is decomposed into 12-channel features by DWT, and concatenated with the 12-channel auxiliary features (output_z) generated by the forward process to form a 24-channel input; after the input is reverse-processed by StegoNet, it is split to obtain 12-channel secret features, which are reconstructed into the restored secret image (rec_secret) by IWT.
[0043] Specifically, the technical solution of this invention includes a forward steganography process and a reverse extraction process. The core architecture consists of a wavelet transform module, a reversible network module (StegoNet), a feature processing module, and a loss function optimization module. These modules work together to ensure the strict reversibility of the "embedding-extraction" process and prevent information loss. The specific technical solution is as follows:
[0044] Step S1: Dataset preprocessing.
[0045] The preprocessing module mainly enhances the dataset and clips its size to meet the model's input requirements. The dataset used is DIV2K. For the test set, random flipping and color perturbation are applied, and finally, it is randomly clipped to a size of 256×256. For the test set, only the center of the data is clipped to 256×256.
[0046] Step S2: Network architecture construction.
[0047] S2-1 Wavelet Transform Module Design:
[0048] This module includes a Discrete Wavelet Transform (DWT) submodule and an Inverse Discrete Wavelet Transform (IWT) submodule, which are used to realize bidirectional conversion between images and frequency domain features, supporting the hierarchical design of "low-frequency covert embedding - high-frequency information carrying".
[0049] The DWT submodule uses a level 1 Haar wavelet transform to decompose the input image (3 channels, 256×256) into 4 subbands: the low-frequency subband (LL) retains the main contour information of the image, with a size of 128×128×3; the high-frequency subbands (LH, HL, HH) correspond to the details and textures in the horizontal, vertical, and diagonal directions, respectively, with a size of 128×128×3; the four subbands are spliced together to obtain 12-channel frequency domain features.
[0050] The IWT submodule, as the inverse operation of DWT, reconstructs the 12-channel frequency domain features into a 3-channel pixel domain image, realizing the conversion of the image from the frequency domain to the spatial domain.
[0051] S2–2, Design of the Reversible Network Module (StegoNet):
[0052] This module consists of 16 cascaded GlowBlocks, enabling reversible fusion of the frequency domain features of the carrier and the secret. The core design ensures "consistency in the number of input-output channels" and "reversibility of the transformation function," providing architectural support for the complete recovery of secret information.
[0053] GlowBlock structure: Each GlowBlock group contains activation normalization (ActNorm), invertible 1×1 convolution (Invertible1x1Conv), and coupling subnet (GlowCouplingSubnet), as detailed below:
[0054] The ActNorm module performs channel-level normalization on the input features. During the initialization phase, it calculates the channel mean using batch data. and standard deviation Set position parameters Scaling parameters Forward transformation Inverse transformation to This ensures that the feature distribution is stable and reversible.
[0055] The reversible 1×1 convolution module initializes the weight matrix as an orthogonal matrix (achieved through QR decomposition). In the forward direction, channel transformation is performed through convolution operations, and in the reverse direction, the inverse matrix is used for reconstruction, ensuring the linear transformation is reversible. Specifically, in the forward flow of the reversible 1×1 convolution module, the weight matrix is initialized as a randomly generated matrix of channel number × channel number dimension, and converted into an orthogonal matrix through QR decomposition, ensuring the reversibility of the forward transformation. In the reverse flow, the inverse transformation matrix is obtained by inverting the weight matrix, and this inverse matrix is reshaped into a convolution kernel shape. The inverse transformation of the input features is achieved through convolution operations. Furthermore, during the floating-point precision conversion of the weight matrix, double-precision calculation of the inverse matrix is used before conversion back to single-precision, ensuring the numerical stability of the inverse transformation.
[0056] The GlowCouplingSubnet serves as a carrier for nonlinear transformation functions and employs a three-level convolutional structure: a 64-channel 3×3 convolution (LeakyReLU activation), a 64-channel 3×3 convolution (LeakyReLU activation), and a 24-channel 3×3 convolution (output scaling factor s and offset t). The parameters of the last convolutional layer are initialized to 0 to ensure that the initial state has minimal impact on the transformation.
[0057] GlowBlock's forward and reverse operations include:
[0058] Forward process: Input 24-channel features, normalize them with ActNorm, and split them into 12-channel x1 and 12-channel x2 by reversible 1×1 convolution. Then, GlowCouplingSubnet calculates s and t for x1. Through the transformation formula, we get y2=e(s)×x2+t (where e(s)=exp(2.0×2×(sigmoid(s)-0.5)), and 2.0 is the clamp parameter). Then, we concatenate x1 and y2 to output 24-channel features.
[0059] Reverse process: Input 24-channel features, split into x1 and y2, then use GlowCouplingSubnet to calculate s and t for x1, restore x2 by x2=(y2-t) / (e(s)+1e-8), concatenate x1 and x2 inputs, perform inverse transformation and ActNorm inverse normalization to obtain the output 24-channel original features.
[0060] S2-3, Feature Processing Module Design:
[0061] This module enables the splicing and splitting of features, supporting the conversion between frequency domain features and pixel domain images.
[0062] Forward stitching: The 12-channel frequency domain features (LL, HL, LH, HH subbands) of the carrier image are stitched together with the 12-channel frequency domain features of the secret image to form a 24-channel input, which is used as the input of the reversible network;
[0063] Reverse splitting: The 24-channel features output by the reversible network are split into 12-channel steganalysis carrier frequency domain features (used to generate steganalysis images) or 12-channel secret frequency domain features (used to recover secret images), and then reconstructed into pixel domain images by IWT.
[0064] Step S3: Forward Steganography Process.
[0065] S3-1, Input of carrier and secret image:
[0066] Input a 256×256×3 cover image and a secret image, and normalize them to the range [0,1] using ToTensor transformation;
[0067] S3-2, Frequency Domain Feature Extraction:
[0068] The carrier image is transformed by wavelet transform to obtain cover_LL, cover_LH, cover_HL, and cover_HH. The secret image is transformed by the same process to obtain secret_LL_dct, secret_LH, secret_HL, and secret_HH.
[0069] S3-3, Feature Fusion and Stereograph Generation:
[0070] The carrier frequency domain features are stitched together to form 12-channel features, and the secret frequency domain features are stitched together to form 12-channel features. The two are superimposed to form 24-channel input features. After the 24-channel features are processed by 16 GlowBlock forward passes, they are split to obtain 12-channel stegographic carrier frequency domain features. The stegographic carrier frequency domain features are reconstructed into a 256×256×3 stegographic image (stego) by the feature reconstruction module (inverse wavelet transform).
[0071] Step S4: Reverse recovery process.
[0072] S4–1, Stegmap Frequency Domain Feature Extraction:
[0073] The steganalysis image (stego) is transformed by wavelet to obtain steggo_LL, steggo_LH, steggo_HL, and steggo_HH, which are then concatenated to form a 12-channel steganalysis frequency domain feature. This feature is then superimposed with a 12-channel noise-filled feature based on a Gaussian distribution to form a 24-channel input feature.
[0074] S4–2, Secret Feature Recovery and Reconstruction:
[0075] After the 24-channel features are processed by 16 GlowBlocks, they are split into 12-channel secret frequency domain features. The secret frequency domain features are reconstructed by the feature reconstruction module (inverse wavelet transform) into the original secret recovery image (rec_secret) of 256×256×3.
[0076] Step S5: Loss calculation and gradient update.
[0077] Loss calculation is performed after each batch of training. This scheme uses a composite loss function, including carrier fidelity loss, information recovery loss, and low-frequency consistency loss, which are set to a priority weight ratio of 3, 3, and 1. After calculating all losses, backpropagation gradient update is performed.
[0078] Step S6: Iterative training.
[0079] After every 10 rounds of training, a model evaluation is performed. The model is then switched to inference mode and tested on the test set. PSNR is used as the main evaluation criterion to verify the model performance. After multiple rounds of iterative training, the model gradually converges.
[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An image steganography method based on frequency domain transform and invertible neural networks, characterized in that, The secret image is written into a carrier image at the sending end and sent to the receiving end. The secret image is reconstructed from the received image at the receiving end. The specific steps include: At the transmitting end, each channel of the 3-channel secret image and carrier image is decomposed into one low-frequency sub-band and three high-frequency sub-bands through wavelet transform; First, all sub-bands of the secret image and the carrier image are spliced together to form 12-channel fusion features. Then, the 12-channel fusion features of the two images are spliced together to form 24-channel fusion features. The 24-channel fused features are input into a reversible network for forward processing to obtain 12-channel stegographic carrier frequency domain features. The frequency domain features of the 12-channel stegographic carrier are used to generate a pixel-domain stegographic image through inverse wavelet transform; At the receiving end, the pixel-domain steganalysis image is decomposed into a low-frequency sub-band and three high-frequency sub-bands by wavelet transform. Each sub-band is superimposed with 12-channel noise-filled features based on Gaussian distribution to form 24-channel input features. The reversible network performs inverse processing on the 24-channel input features, splitting the 24-channel input features into 12-channel secret frequency domain features; The 12-channel secret frequency domain features are reconstructed into a secret restored image through inverse wavelet transform.
2. The image steganography method based on frequency domain transformation and reversible neural networks according to claim 1, characterized in that, In the process of decomposing the image into one low-frequency sub-band and three high-frequency sub-bands by wavelet transform, the first-order Haar wavelet transform is used to decompose the image. The low-frequency sub-band retains the main contour information of the image, while the three high-frequency sub-bands correspond to the detailed textures in the horizontal, vertical and diagonal directions, respectively.
3. The image steganography method based on frequency domain transformation and reversible neural networks according to claim 1, characterized in that, The reversible network consists of 16 cascaded GlowBlocks. Each GlowBlock includes a cascaded normalization module, a reversible 1×1 convolution module, and a coupling layer subnet. Specifically, the normalization module performs channel-level normalization on the input image, the reversible 1×1 convolution module splits the normalized data into channels, and the coupling layer subnet calculates the scaling factor and offset based on the first feature map obtained from the split. It then performs a transformation based on the scaling factor, offset, and the second feature map obtained from the split. The transformed feature map is concatenated with the first feature map and used as the output of the coupling layer subnet.
4. The image steganography method based on frequency domain transformation and reversible neural networks according to claim 3, characterized in that, In the normalization module, position parameters are constructed based on the mean and standard deviation of each channel pixel. and scaling parameters In the forward processing, normalization is represented as In the reverse processing, normalization is represented as ,in For the normalized data, This is the data before normalization.
5. The image steganography method based on frequency domain transformation and reversible neural network according to claim 4, characterized in that, Position parameters Represented as Scaling parameters Represented as , This is the mean of pixel values at all spatial locations across all batch samples in a single channel of the input image. This represents the variance of pixel values at all spatial locations across all batch samples in a single channel of the input image.
6. The image steganography method based on frequency domain transformation and invertible neural networks according to claim 3, characterized in that, The reversible 1×1 convolution module initializes the weights with an orthogonal matrix and performs a 1×1 convolution of the channel dimension on the spliced frequency domain tensor after discrete wavelet transform, realizing reversible channel shuffling of frequency domain features; the module completes the inverse operation through the transpose of the orthogonal matrix.
7. The image steganography method based on frequency domain transformation and reversible neural networks according to claim 3, characterized in that, The specific data processing procedures of the coupled layer subnet include: During the forward flow, the input to the coupled layer subnet is the first image feature x1 and the second image feature x2. The first image feature x1 is processed through a three-level convolutional structure. The first level uses a 64-channel 3×3 convolution operation and a LeakyReLU activation operation. The second level uses a 64-channel 3×3 convolution operation and a LeakyReLU activation operation. The third level uses a 24-channel 3×3 convolution. The three-level convolutional structure outputs a scaling factor s and an offset t. The second image feature x2 is transformed using the scaling factor s and the offset t to obtain the transformed image feature, i.e., y2=e(s)×x2+t. The function e(s) is expressed as e(s)=exp(2.0×2×(sigmoid(s)-0.5)). The transformed image feature y2 is concatenated with the first image feature x1 as the output of the coupled layer subnet in the forward flow. In the reverse process, the input to the coupled layer subnet is the first image feature x1 and the transformed image feature y2. The first image feature x1 is processed through a three-level convolutional structure to calculate the scaling factor s and the offset t. The transformed image feature y2 is then restored using the scaling factor s and the offset t to obtain the second image feature x2. The first image features and the second image features are concatenated together as the output of the coupling layer subnet in the reverse process.
8. The image steganography method based on frequency domain transformation and reversible neural network according to claim 1, characterized in that, During the training of the reversible network, the loss function includes three types of sub-losses: carrier steganalysis loss, secret reconstruction loss, and low-frequency loss. The steganalysis loss constrains the pixel similarity between the stegana image and the original carrier image. The secret reconstruction loss constrains the pixel consistency between the back-extracted secret image and the original secret image. The low-frequency loss constrains the similarity between the high-frequency subbands of the stegana image after wavelet decomposition and the high-frequency subbands of the original carrier image after wavelet decomposition.
9. The image steganography method based on frequency domain transformation and reversible neural network according to claim 8, characterized in that, The weights for carrier steganography loss, secret reconstruction loss, and low-frequency loss are 3, 3, and 1, respectively.
10. An image steganography system based on frequency domain transform and reversible neural networks, characterized in that, To implement the image steganography method based on frequency domain transformation and reversible neural network as described in claim 1, the method includes: The wavelet transform module is used to perform frequency domain decomposition of the carrier image, secret image, and stegana. The inverse wavelet transform module is used to reconstruct the image from frequency domain features to pixel domain; The reversible network consists of 16 GlowBlocks connected in series. Each GlowBlock includes a cascaded activation normalization layer, a reversible 1×1 convolutional layer, and an affine coupling layer, which are used to achieve reversible transformation of 24-channel fused features.