Image steganography method based on residual improved MWCNN network
By improving the MWCNN network, employing wavelet transform and residual structure, and combining a hybrid loss function and Adam optimization algorithm, the problems of low hiding capacity and extraction quality in color image steganography schemes are solved, achieving more efficient image steganography results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGNAN UNIV
- Filing Date
- 2023-07-04
- Publication Date
- 2026-07-24
AI Technical Summary
Existing color image steganography schemes suffer from limited hiding capacity, low security of the hidden image, and poor quality of extracted image. Furthermore, deep learning image steganography schemes are limited in the number of network layers, resulting in poor image steganography performance.
We employ a residual-improved MWCNN network, replacing pooling operations with discrete wavelet transform, and combining residual structure and wavelet component skip-layer fusion to design a symmetric network for image hiding and extraction. We then use a hybrid loss function and the warm-start Adam optimization algorithm to optimize the network parameters.
It improves the hiding capacity and security of image steganography, while enhancing the extraction quality and accuracy of secret images, thus achieving better image steganography results.
Smart Images
Figure CN116912074B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an image steganography method based on residual improved MWCNN network, belonging to the field of covert communication technology. Background Technology
[0002] In the information age, images, with their intuitive and vivid advantages, have gradually become a primary medium for information exchange. However, images containing sensitive information are vulnerable to attacks during transmission, leading to the leakage of private data. Therefore, ensuring the security of image transmission is essential. Image encryption can protect content security, but it can easily attract attention, resulting in the interception or destruction of information; while image steganography, as a branch of covert communication, can achieve the covert transmission of secret images.
[0003] Image steganography works by hiding secret information in redundant locations within a carrier image without causing significant distortion. It includes spatial domain steganography and transform domain steganography. Spatial domain steganography involves slightly modifying the pixel intensity values of the host image based on hidden secret data. Transform domain steganography, on the other hand, inserts the secret data into the spectral components of certain transforms, utilizing various image transforms such as Fourier transforms and wavelet transforms. Relatively speaking, transform domain steganography offers higher security and robustness than spatial domain steganography because modifications to image pixels are more easily detected. Furthermore, transform domain steganography is less affected by rotation, scaling, and compression.
[0004] To address the limitations of current color image steganography schemes, such as limited hiding capacity, low security of the hidden image, and poor extraction quality, existing methods combine deep learning with steganography, utilizing the data-driven approach and self-learning capabilities of convolutional neural networks for image hiding and extraction. Furthermore, to minimize information loss during transformation, many deep learning image steganography schemes omit pooling layers in the network. However, this increases the number of parameters, limiting the number of network layers and consequently leading to poor image steganography results. Subsequently, some researchers proposed using the UNet network for image steganography; however, the downsampling of the UNet network causes irreversible information loss, thus limiting the steganography capacity and the extraction quality of the hidden image. Summary of the Invention
[0005] To address the issues of limited hiding capacity, low security of encrypted images, and low quality of extracted images, this invention provides an image steganography method based on a residual improved MWCNN network, comprising:
[0006] Step 1: Preprocess the secret image to the same size as the carrier image, and then stitch the secret image and the carrier image of the same size together by channel concatenation;
[0007] Step 2: Input the spliced secret and carrier images into the residual improved MWCNN network for image hiding. The residual improved MWCNN network is a symmetrical network. The input image first passes through the contraction path on the left side of the network, and then through the expansion path on the right side of the network to obtain the corresponding secret image.
[0008] Step 3: Input the secret image into the extraction network consisting of convolutional layers, activation layers, and DWT to extract the secret image;
[0009] The improved MWCNN network and the extraction network are the networks after training. During training, the loss values of the residual improved MWCNN network and the extraction network are calculated using a loss function, which is a hybrid loss function obtained by combining the loss function of the residual improved MWCNN network and the loss function of the extraction network. During training, the Adam optimization algorithm with warm start is used to optimize the network parameters. At the same time, batch normalization and Leaky ReLU activation function are added to the network to speed up the network training.
[0010] Optionally, step 1 includes:
[0011] The secret image and the carrier image are cropped to obtain secret images and carrier images of the same size;
[0012] Using equation (1), the secret image and the carrier image are stitched together by channel number;
[0013] X = concat[H,S], H, S∈R L×W×N Equation (1)
[0014] In equation (1), H represents the carrier image, S represents the secret image, and X is a tensor of size L×W×(2×N) obtained after preprocessing. The spliced tensor is used as the input of the residual improved MWCNN network.
[0015] Optionally, step 2 includes:
[0016] The shrinking path of the hidden network is used to first perform a 3×3 convolution on the input, and then the convolution output passes through the activation layer and the batch normalization layer, as shown in Equation (2):
[0017]
[0018] Where, suppose X i-1 If X is the input to the convolutional layer, then... i For the corresponding output features; w i These are the weights corresponding to the convolution operation; b represents the convolution operation; iThis represents the threshold of the layer; after convolution, the f(·) activation function is used to achieve non-linear mapping;
[0019] Then, equation (3) uses discrete wavelet transform to downsample the feature map and compress the image size:
[0020] X i+1 =DWT(X i Equation (3)
[0021] In the formula, DWT represents Discrete Wavelet Transform, which transforms X... i X is obtained by downsampling using discrete wavelet transform. i+1 ;
[0022] Then, residual learning is performed on the feature map using equation (4):
[0023] X i+2 =X i+1 +C(X i+1 Equation (4)
[0024] In the formula, C(X) i+1 ) for X i+1 The residual between the feature map after two convolution operations and the input;
[0025] The network expansion path adopts a similar structure to the contraction path, but unlike the contraction path, the expansion path uses the inverse discrete wavelet transform shown in equation (5) to achieve image upsampling:
[0026] X' i =IDWT(X' i+1 Equation (5)
[0027] Wherein, IDWT represents the inverse discrete wavelet transform;
[0028] During the feature fusion process, the different detail components obtained by downsampling in the shrinking path are increased in the number of feature channels by 1×1 convolution, and then added to the feature map of the expanding path by element-wise summation.
[0029] Repeat the above steps until the output of the hidden network, L×W×N, is obtained after three rounds of downsampling and upsampling.
[0030] Optionally, step 3 includes:
[0031] An extraction network is constructed using convolutional layers, activation layers, and DWT.
[0032] The output of the residual improved MWCNN in step 2 is used as the input of the extraction network. After 9 rounds of convolution, two rounds of downsampling and upsampling, the final extracted image is obtained. The sampling process also uses discrete wavelet transform and inverse transform to transform the extraction process to the wavelet domain.
[0033] Optionally, the method for determining the loss function of the residual-improved MWCNN network and the extraction network includes:
[0034] The difference between the hidden image and the carrier image, and between the secret image and the extracted image, is calculated using mean squared error. Weight parameters are set to obtain the final hidden network, and the network loss is extracted.
[0035]
[0036] In the formula, H and HS represent the carrier image and the hidden image output by the hidden network, respectively; S and S' represent the secret image and the extracted image output by the extraction network, respectively; L1 represents the loss function of the hidden network, L2 represents the loss function of the extraction network, and L... total Let l1 and l2 represent the mixed loss function, and l1 and l2 represent the corresponding weights, respectively.
[0037] Optionally, when using the Adam optimization algorithm with warm start to optimize network parameters, a learning rate lower than the initial Adam learning rate is set in the early stage of training, and the learning rate is continuously increased to the set initial learning rate as the number of iterations increases.
[0038] Optionally, the learning rate is set to 0.0002 in the initial stage of training.
[0039] Optionally, the secret image and the carrier image are RGB color images.
[0040] This application also provides a covert communication method, in which the above-mentioned image steganography method is used to hide and extract images during the covert transmission of secret images.
[0041] This application also provides the application of the above-mentioned image steganography method in covert communication.
[0042] The beneficial effects of this invention are:
[0043] By improving the image steganography scheme based on MWCNN network, the network downsampling is changed from the traditional pooling operation to wavelet transform. The residual network and wavelet component skip-layer fusion are introduced into the image steganography network, which can achieve better hiding effect in color image steganography. This method can also improve the visual quality of the extracted image. This steganography scheme can hide a large amount of image information while ensuring the quality of the encrypted image and the accuracy of the extracted image. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a structural diagram of the image hiding network, which is the residual improved MWCNN network obtained by residual improvement of MWCNN.
[0046] Figure 2 This is a diagram of the image extraction network structure.
[0047] Figure 3 This is a simulation diagram of the image hiding and extraction results under the following conditions: the training and test sets are miniImageNet, the input image is a color image with a size of 256×256×3, the optimization function is Adam, the loss function weights are l1=1, l2=0.75, the initial learning rate of the network is 0.0002, and the number of training rounds is 300. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0049] Multi-Level Wavelet Convolutional Neural Networks (MWCNNs) are multi-level wavelet CNNs that use Discrete Wavelet Transform (DWT) instead of pooling operations. They are commonly used in image restoration and image classification. However, if directly applied to image steganography, the increased number of network layers may make the network difficult to train, and element-wise summation-based skip connections may lead to image information loss, thus affecting the image steganography effect. Therefore, this application improves upon MWCNNs. First, it replaces pooling operations with Discrete Wavelet Transform, using wavelet transform and inverse transform to achieve downsampling and upsampling of the image hiding network. Furthermore, it improves upon MWCNNs by replacing element-wise summation-based skip connections with feature channel concatenation to reduce information loss during feature fusion. Next, to supplement image details, different detail components (high-frequency components) of the wavelet transform are added to the feature map of the extended path during feature fusion. Finally, a residual structure is used in the hiding network to prevent network degradation caused by the increase in the number of network layers, while simultaneously optimizing network training, thereby improving the image hiding effect.
[0050] Based on the size relationship between the secret image to be hidden and the carrier image used for hiding, image hiding can be divided into large-capacity hiding and small-capacity hiding. In small-capacity hiding, the secret image is smaller than the carrier image, while in large-capacity hiding, the secret image is the same size as or larger than the carrier image. Therefore, small-capacity hiding is easier to implement. To highlight the effectiveness of the method in this application, subsequent embodiments will be described using the example of the secret image and carrier image being the same size:
[0051] Example 1:
[0052] This embodiment provides an image steganography method based on a residual improved MWCNN network, the method comprising:
[0053] Step 1: Preprocess the secret image to the same size as the carrier image, and then stitch the secret image and the carrier image of the same size together by channel stitching;
[0054] Step 2: Input the spliced secret and carrier images into the residual improved MWCNN network for image hiding. The residual improved MWCNN network is a symmetrical network. The input image first passes through the contraction path on the left side of the network, and then through the expansion path on the right side of the network to obtain the corresponding secret image.
[0055] Step 3: Input the secret image into the extraction network consisting of convolutional layers, activation layers, and DWT to extract the secret image;
[0056] The improved MWCNN network and the extraction network are the networks after training. During training, the loss values of the residual improved MWCNN network and the extraction network are calculated using a loss function, which is a hybrid loss function obtained by combining the loss function of the residual improved MWCNN network and the loss function of the extraction network. During training, the Adam optimization algorithm with warm start is used to optimize the network parameters. At the same time, batch normalization and Leaky ReLU activation function are added to the network to speed up the network training.
[0057] Example 2
[0058] This embodiment provides an image steganography method based on a residual improved MWCNN network, applicable to RGB color images. The method includes:
[0059] Step 1: Crop the secret image and the carrier image to obtain secret images and carrier images of the same size;
[0060] Using equation (1), the secret image and the carrier image are stitched together by channel number;
[0061] X = concat[H,S], H, S∈R L×W×N Equation (1)
[0062] In equation (1), H represents the carrier image, S represents the secret image, and X is a tensor of size L×W×(2×N) obtained after preprocessing.
[0063] Step 2: Input the stitched secret and carrier images into the residual improved MWCNN network for image hiding to obtain the corresponding secret image;
[0064] Design as Figure 1 The network shown performs image hiding on the input image. The left-hand network is the shrinking path, whose input is an L×W×(2×N) tensor, and the right-hand network is the expanding path.
[0065] The network shrinkage path first performs a 3×3 convolution (Conv) on the input, then passes it through an activation layer (Leaky ReLU) and a batch normalization layer (BN), as shown in equation (2):
[0066]
[0067] Where, suppose X i-1 If X is the input to the convolutional layer, then... i For the corresponding output features; w i These are the weights corresponding to the convolution operation; b represents the convolution operation; i This represents the threshold of the layer; after convolution, the f(·) activation function is used to achieve non-linear mapping;
[0068] Then, equation (3) uses discrete wavelet transform to downsample the feature map and compress the image size:
[0069] X i+1 =DWT(X i Equation (3)
[0070] In the formula, DWT represents Discrete Wavelet Transform, which transforms X... i X is obtained by downsampling using discrete wavelet transform. i+1 .
[0071] Then, residual learning is performed on the feature map using equation (4):
[0072] X i+2 =X i+1 +C(X i+1 Equation (4)
[0073] In the formula, C(X) i+1 ) for X i+1 The residual between the feature map after two convolution operations and the input.
[0074] The network expansion path adopts a similar structure to the contraction path, but unlike the contraction path, the expansion path uses the inverse discrete wavelet transform shown in equation (5) to achieve image upsampling:
[0075] X' i =IDWT(X' i+1 Equation (5)
[0076] Meanwhile, in order to supplement image details, during the feature fusion process, the different detail components obtained by downsampling in the shrinking path are increased in the number of feature channels by 1×1 convolution, and then added to the feature map of the expanding path by element-wise summation.
[0077] Repeat the above steps until the output of the hidden network, L×W×N, is obtained after three rounds of downsampling and upsampling.
[0078] Step 3: Input the secret image into the extraction network consisting of convolutional layers, activation layers, and DWT to extract the secret image;
[0079] Design as Figure 2 The network shown extracts secret images from the input image. Input is the output secret image obtained in step 3, which undergoes nine rounds of convolution, two rounds of downsampling and upsampling to obtain the final extracted image. The sampling process also uses discrete wavelet transform and inverse transform to convert the extraction process to the wavelet domain.
[0080] Step 4: The network adopts the loss function design principle of autoencoders, combining the hidden network loss function and the extraction network loss function to obtain a hybrid loss function; secondly, the Adam optimization algorithm with warm start is used to optimize the network parameters during training, and batch normalization and Leaky ReLU activation function are added to the network to speed up the network training.
[0081] Equation (6) uses mean square error to calculate the difference between the hidden image and the carrier image, and between the secret image and the extracted image, and sets weight parameters to obtain the final loss of the hidden network and the extraction network.
[0082]
[0083] In the formula, H and HS are the carrier image and the hidden image output by the hiding network, respectively; S and S' are the secret image and the extracted image output by the extraction network, respectively.
[0084] Subsequently, in order to obtain a suitable learning rate to improve the learning efficiency of the network, the Adam optimization algorithm with warm start was adopted. In the early stage of training, a learning rate lower than the initial Adam learning rate was set. As the number of iterations increased, the learning rate was continuously increased to the set initial learning rate.
[0085] Adam uses second-order momentum to obtain an adaptive learning rate, and the process is shown in Algorithm 1:
[0086]
[0087] Where, m t For g t The first moment, i.e., g t Expectations; e t For g t The second moment, i.e., g t 2 The expected value; β1 and β2 are the exponential decay coefficients of the first and second moments, respectively. and m t e t The bias correction; τ is a constant, usually set to 10. -8 .
[0088] To make the objectives, technical solutions, and advantages of this invention clearer, the following comparison will be made between some classic image steganography schemes and the proposed image steganography scheme, demonstrating the superiority of the image steganography scheme based on residual improved MWCNN network in terms of visual security in hiding secret images and image extraction quality.
[0089] The image steganography used for simulation is the image hiding and extraction effect of the Baluja steganography scheme and the U-Net structure steganography scheme.
[0090] Among them, the Baluja steganography scheme is a classic image steganography scheme. It was the first to propose using deep neural networks for image steganography in image steganography tasks, and it showed good hiding effect. The U-Net structure steganography scheme adopts a structure similar to the U-Net network as the basic architecture of the hiding network. It combines the feature maps of the shrinking path and the expanding path through skip connections, so as to achieve better hiding and extraction effects.
[0091] Experimental results are as follows Figure 3 As shown in Tables 1 and 2, Figure 3 The dataset contains four columns of images: the first column is the original carrier image, the second column is the carrier image output by the hiding network, the third column is the original secret image, and the last column is the extracted image output by the extraction network. Table 1 shows the PSNR and MSSIM values for image hiding and extraction under the following conditions: training and testing sets are miniImageNet, input images are color images with a size of 256×256×3, the optimization function is Adam, the loss function weights are l1=1, l2=0.75, the initial learning rate of the network is 0.0002, and the number of training epochs is 300. The eight sets of data in Table 1 correspond to... Figure 3The eight sets of images in the text.
[0092] Table 2 compares the image hiding and extraction PSNR values of the Baluja steganography scheme, the U-Net structure steganography scheme, and the image steganography scheme based on the residual improved MWCNN network of this application under the following conditions: training and test sets are miniImageNet, input images are color images with a size of 256×256×3, the optimization function is Adam, the loss function weights are l1=1, l2=0.75, the initial learning rate of the network is 0.0002, and the number of training epochs is 300.
[0093] The Baluja steganography scheme can be found in "Baluja S. Hiding images in plain sight: Deepsteganography[J]. Advances in neural information processing systems, 2017:2069-2079". This method uses a fully convolutional neural network for image steganography, and the network includes a cryptographic image preprocessing network, a secret image hiding network, and a secret image extraction network. The U-Net structure steganography scheme can be found in "Duan X, Jia K, Li B, et al. Reversible image steganography scheme based on a U-Net structure[J]. IEEE Access, 2019, 7:9314-9323". This method uses a structure similar to the U-Net network as the secret image hiding network. MWCNN uses the original MWCNN network as the hiding network and convolutional layers and activation layers as the extraction network. The residual improved MWCNN uses the residual improved MWCNN as the hiding network and convolutional layers and activation layers as the extraction network.
[0094] Table 1: PSNR and MSSIM values of steganographic images on the test set
[0095]
[0096] Table 2: Comparison of Steganography Effects of Different Schemes
[0097]
[0098] from Figure 3As shown in Table 1, the steganography scheme of the present invention can effectively achieve secret image hiding and extraction. From the comparison of the image hiding and extraction PSNR values of the Baluja steganography scheme, the U-Net structure steganography scheme, and the image steganography scheme based on the residual improved MWCNN network in Table 2, it is found that the steganography scheme of the present invention can effectively improve the PSNR value of image hiding and extraction.
[0099] Figure 3 This invention provides image hiding and extraction results under the following conditions: training and testing sets are miniImageNet; input images are color images with a size of 256×256×3; the optimization function is Adam; the loss function weights are l1=1, l2=0.75; the initial learning rate of the network is 0.0002; and the number of training epochs is 300. Figure 3 It can be seen that the image after steganography using the steganography method of this application is almost indistinguishable from the original image, and the extracted result is almost indistinguishable from the secret image. Based on Table 2, it can be seen that the steganography scheme of this invention can effectively improve the PSNR value of image hiding and extraction.
[0100] In summary, under the same training and test sets, the residual-improved MWCNN network image steganography scheme based on the present invention can achieve more secure image hiding and secret image extraction effects.
[0101] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.
[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An image steganography method based on residual improved MWCNN network, characterized in that, The method includes: Step 1: Preprocess the secret image to the same size as the carrier image, and then stitch the secret image and the carrier image of the same size together by channel concatenation; Step 2: Input the spliced secret and carrier images into the residual improved MWCNN network for image hiding. The residual improved MWCNN network is a symmetrical network. The input image first passes through the contraction path on the left side of the network, and then through the expansion path on the right side of the network to obtain the corresponding secret image. Step 3: Input the secret image into the extraction network consisting of convolutional layers, activation layers, and DWT to extract the secret image; The improved MWCNN network and the extraction network are the networks after training. During training, the loss values of the residual improved MWCNN network and the extraction network are calculated using a loss function, which is a hybrid loss function obtained by combining the loss function of the residual improved MWCNN network and the loss function of the extraction network. During training, the Adam optimization algorithm with warm start is used to optimize the network parameters. At the same time, batch normalization and Leaky ReLU activation function are added to the network to speed up the network training. Step 2 includes: Residual improvements are applied to MWCNN to obtain a residual-improved MWCNN network, which is then used as a hidden network. The input is first processed using the shrinking path of the hidden network. Convolution, followed by the convolution output passing through an activation layer and a batch normalization layer, as shown in equation (2): Equation (2) Where, assuming If the input is the convolutional layer, then The corresponding output features; These are the weights corresponding to the convolution operation; This represents the convolution operation; This represents the threshold of the layer; it is used after convolution. Activation functions implement nonlinear mappings; Then, equation (3) uses discrete wavelet transform to downsample the feature map and compress the image size: Equation (3) In the formula, DWT represents Discrete Wavelet Transform, which... The result is obtained by downsampling using discrete wavelet transform. ; Then, residual learning is performed on the feature map using equation (4): Equation (4) In the formula, To The residual between the feature map after two convolution operations and the input; The network expansion path adopts a similar structure to the contraction path, but unlike the contraction path, the expansion path uses the inverse discrete wavelet transform shown in equation (5) to achieve image upsampling: Equation (5) Wherein, IDWT represents the inverse discrete wavelet transform; During feature fusion, the different detail components obtained by downsampling in the shrinking path are increased in the number of feature channels by 1×1 convolution, and then added to the feature map of the expanding path by element-wise summation. Based on the MWCNN network, the skip connection based on element-wise summation is changed to feature channel concatenation to reduce information loss during feature fusion. Repeat the above steps until the output of the hidden network is obtained after three rounds of downsampling and upsampling. Output.
2. The method according to claim 1, characterized in that, Step 1 includes: The secret image and the carrier image are cropped to obtain secret images and carrier images of the same size; Using equation (1), the secret image and the carrier image are spliced together by channel number; Equation (1) In equation (1), H Represents a carrier image. S Represents a secret image. X The size obtained after preprocessing is The tensors are concatenated and used as the input of the residual improved MWCNN network.
3. The method according to claim 2, characterized in that, Step 3 includes: An extraction network is constructed using convolutional layers, activation layers, and DWT. The output of the residual improved MWCNN in step 2 is used as the input of the extraction network. After 9 rounds of convolution, two rounds of downsampling and upsampling, the final extracted image is obtained. The sampling process also uses discrete wavelet transform and inverse transform to transform the extraction process to the wavelet domain.
4. The method according to claim 3, characterized in that, The methods for determining the loss functions of the residual improved MWCNN network and the extraction network include: The difference between the hidden image and the carrier image, and between the secret image and the extracted image, is calculated using mean squared error. Weight parameters are set to obtain the final hidden network, and the network loss is extracted. Equation (6) In the formula, These are the carrier image and the hidden image output by the hidden network, respectively. These are the secret image and the extracted image output by the extraction network, respectively. L 1 indicates the hidden network loss function. L 2 indicates the extraction of the network loss function. L total Represents the mixed loss function. l 1 and l 2 represents the corresponding weights.
5. The method according to claim 4, characterized in that, When using the Adam optimization algorithm with warm start to optimize network parameters, a learning rate lower than the initial Adam learning rate is set in the early stage of training. As the number of iterations increases, the learning rate is continuously increased to the set initial learning rate.
6. The method according to claim 5, characterized in that, The learning rate was set to 0.0002 in the initial stage of training.
7. The method according to claim 6, characterized in that, The secret image and the carrier image are RGB color images.
8. A covert communication method, characterized in that, The method employs the image steganography method described in any one of claims 1-7 to hide and extract images during the covert transmission of secret images.
9. The image steganography method according to any one of claims 1-7 in covert communication.