CN114897658A - Image steganography model protection method based on extraction network parameter scrambling - Google Patents

Abstract

The invention provides an image steganography model protection method based on extraction network parameter scrambling, and belongs to the technical field of image encryption. Since the sender may be stolen by a third party when sending the extracting network to the receiver for the first time, the secret image of the subsequent transmission is leaked. For this, the sender scrambles the parameters of the extracted network model by a scrambling algorithm agreed with the receiver, and then reconstructs a new extracted network model by extracting the parameters of the network model after scrambling. When the sender sends the new extraction network and the secret-carrying image to the receiver, even if the extraction network model is acquired by a third party in the transmission process, the similarity between the secret image extracted from the secret-carrying image and the original secret image is low, so that the secret image is prevented from being leaked. And the receiver can restore the new extraction network model to the extraction network model before scrambling according to the agreed scrambling algorithm, and extract the secret image from the secret image, thereby enhancing the safety of the image steganography process.

Description

An image steganographic model protection method based on extracting network parameters and scrambling

technical field

The invention relates to an image steganographic model protection method based on extracting network parameter scrambling, and belongs to the technical field of image encryption.

Background technique

In daily work, we cannot do without computers to transmit information, because computers can meet people's daily life and work needs. In order to ensure that secret information can be safely transmitted in the network, people usually choose encryption technology to protect information, such as file encryption, image encryption, encrypted files have an obvious disadvantage, encrypted files cannot be opened normally, so encryption The post file can easily arouse suspicion from others during the transfer process. Information hiding can solve the above problems very well. Information hiding refers to hiding the secret information we want to transmit in the media that can be disclosed, and it is difficult for people to find the existence of secret information with intuitive hearing and vision. There are many carriers used for information hiding, such as images, sounds, videos, documents, etc.

Digital images are one of the most common carriers of information hiding. There are three main reasons for choosing digital images as carriers: (1) In the context of the information technology era, there are a lot of picture information on the Internet, and people can easily obtain carriers; (2) ) The semantic content contained in the image is very rich and can express different information; (3) Compared with text and audio, digital images have larger storage capacity and greater data redundancy, and are very suitable as carriers for information hiding , hence the emergence of image steganography. Image steganography technology means that the sender embeds the secret information to be transmitted into the carrier image through certain technical means, and then sends the secret image containing the secret image to the receiver through an open network channel, and the receiver receives the secret image. After the image, the secret information is extracted from the carrier image through decryption technology, so as to avoid the third party's detection.

Image steganography roughly goes through three stages. In the first stage, researchers implement image steganography by modifying the pixel values of the image. For example, LSB steganography, which achieves the purpose of information hiding by modifying or replacing the least significant bits. This method will not cause much impact on the overall visual effect of the image, but it is relatively easy to be detected by steganalysis technology, and LSB steganography technology is difficult to meet the requirements of modern communication security. In the second stage, the researchers consider that the image has a high degree of complexity and more redundancy, so the content of the image is modified by the content-adaptive steganography method to achieve information hiding, which makes it visually difficult for people to perceive and difficult to detect. Anomalies are detected by steganalysis techniques. The more commonly used adaptive steganography methods include HUGO, WOW, UNIWARD, HILL, etc., but with the further improvement of steganalysis technology, SRM can analyze and detect adaptive steganography. In the third stage, with the rise of Deep Convolutional Neural Networks (DCNN) in various fields, researchers introduced DCNN into the field of image steganography to achieve better image steganography. For example, the encoder-decoder image steganography framework proposed by Baluja et al., including image preprocessing stage, hidden network, and extraction network, can not only ensure that the secret image and the carrier image are consistent, but also the extracted secret image and the original secret image. There is almost no distortion visually, and it has a certain ability to resist steganalysis.

The image steganography network model based on deep learning is mainly composed of a hidden network and an extraction network. First, the sender selects its own image data set for training to obtain an image steganography network model, and then the sender and receiver pass the trained hidden network model. Write network models for covert communication. When the sender and the receiver communicate for the first time, the sender needs to send the parameters of the trained extraction network model to the receiver, thus establishing the preparation for covert communication. The receiver can only get the parameters of the extraction network model. The extraction network model is established, and then the secret information hidden in the secret image is extracted through the extraction network model. However, there is a certain risk of leakage when the two parties extract the parameters of the network model for the first transmission, and the security level is low. If the third party intercepts the transmission process, the subsequent communication content may be obtained by the third party through the extraction network. .

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image steganographic model protection method based on scrambling extracted network parameters, which is used to solve the problem of the security level when the sender sends the extracted network model in the image steganography network to the receiver in the image steganography method. lower problem.

In order to achieve the above object, the present invention provides an image steganographic model protection method based on extracting network parameters and scrambling, comprising the following steps:

S1. The sender hides the secret image to be sent into the carrier image through the trained hidden network model, obtains the corresponding secret image, and obtains the trained extraction network model. The hidden network model and the extraction network model together constitute Deep learning image steganography network;

S2. The sender performs a scramble operation on the parameters of the trained extraction network model through the agreed scrambling algorithm, and reconstructs the parameters of the scrambled extraction network model to obtain a new extraction network model; the sender will The secret image and the new extraction network model are sent to the receiver; the agreed scrambling algorithm satisfies: for the same secret image, the similarity of the secret image extracted by the extraction network model before scrambling and the extraction network model after scrambled below the set threshold;

S3. After receiving the new extraction network model, the receiver scrambles and decrypts the new extraction network model according to the agreed scrambling algorithm to obtain the extraction network model before scrambling, and according to the extraction network model before scrambling The received secret image is extracted to obtain the secret image.

When the sender sends the extraction network to the receiver for the first time, it may be stolen by a third party, resulting in the disclosure of the secret images of subsequent transmissions. In this regard, the sender first scrambles the parameters of the extracted network model through a scrambling algorithm agreed with the receiver, and then reconstructs a new extracted network model by scrambling the parameters of the extracted network model. When the sender sends the new extraction network and the secret image to the receiver, even if the extraction network model is acquired by a third party during the transmission process, the secret image extracted from the secret image is still larger than the original secret image. gap, the similarity is low, thus avoiding the leakage of secret images. The receiver can restore the new extraction network model to the extraction network model before scrambling according to the agreed scrambling algorithm, and extract the secret image to obtain an effective secret image, which enhances the security of the image steganography process.

Further, in the above method, both the hidden network model and the extraction network model use a deep neural convolutional network, and the scrambling operation refers to performing a scrambling operation on the positions of the convolution kernels in one or more convolutional layers in the trained extraction network model. Inter-layer scrambling, or inter-layer scrambling for the positions of convolution kernels in different convolutional layers; when intra-layer scrambling occurs, the position of the convolution kernel only changes within the convolution layer, and the position of the convolution kernel when inter-layer scrambling occurs Changes between different convolutional layers.

The extraction network model adopts a deep neural convolutional network, including multiple convolutional layers. During the scrambling operation, perform intra-layer scrambling or inter-layer scrambling on the trained extraction network model. The method for intra-layer scrambling is to change the position of the convolution kernel in the convolution layer; the method for inter-layer scrambling is: : Change the positions of the convolution kernels of different convolutional layers. By changing the position of the convolution layer, the sequence of convolution operations after the encrypted image is input to the extraction network model is changed, thereby realizing the encryption of the extraction network model and enhancing the security of the image steganography process.

Further, in the above method, the agreed scrambling algorithm adopts a super Lorenz chaotic system.

For the agreed scrambling algorithm, a specific implementation method is provided. The sender and the receiver mutually agree on the control parameters of the super-Lorenz chaotic system and the process of the above scrambling operation, so as to realize the agreement of the scrambling algorithm.

Further, in the above method, the structural similarity analysis method SSIM is used to determine the similarity.

For the same secret image, it is necessary to analyze the similarity between the extraction network model before scrambling and the secret image extracted by the extraction network model after scrambling, and provide a specific method to determine the similarity, which is convenient for the implementation of the present invention .

Further, in the above method, the extraction network model includes an upsampling module, a pooling layer, a downsampling module and a convolutional layer that are connected to each other; the upsampling module and the downsampling module each include 4 convolutional layers.

A specific extraction network model is provided to facilitate the implementation of the present invention.

Further, in the above method, when the scrambling operation is performed in step S2, in-layer scrambling is performed on the first, second or fourth convolutional layer in the upsampling module, or the third convolutional layer in the downsampling module is scrambled. 1st or 4th convolutional layer for intra-layer scrambling.

For the extraction network model mentioned above, by performing intra-layer scrambling on the 1st, 2nd or 4th convolutional layer in the upsampling module, or by scrambling the 3rd or 4th convolutional layer in the downsampling module By layer-by-layer chaos, a new extraction network model with better effect can be reconstructed, and the operation is simple.

Further, in the above method, when the scrambling operation is performed in step S2, the first, second and fourth convolutional layers in the up-sampling module, and the third and fourth in the down-sampling module are The convolutional layers perform inter-layer scrambling.

For the extraction network model mentioned above, the first, second and fourth convolutional layers in the upsampling module and the third and fourth convolutional layers in the downsampling module are inter-layered. A new extraction network model with better effect can be reconstructed, which is convenient for the implementation of the present invention and has a high degree of security.

Further, in the above method, when the scrambling operation is performed in step S2, inter-layer scrambling is performed on the four convolutional layers of the upsampling module.

For the extraction network model mentioned above, a new extraction network model with better effect can be reconstructed by performing inter-layer scrambling on the four convolutional layers of the upsampling module, which is convenient for the implementation of the present invention, and has a high degree of security. high.

Further, in the above method, when the scrambling operation is performed in step S2, inter-layer scrambling is performed on the four convolutional layers of the downsampling module.

For the extraction network model mentioned above, a new extraction network model with better effect can be reconstructed by performing inter-layer scrambling on the four convolutional layers of the downsampling module, which is convenient for the implementation of the present invention, and has a high degree of security. high.

Further, in the above method, when the scrambling operation is performed in step S2, inter-layer scrambling is performed on the 9 convolutional layers in the extraction network model.

For the extraction network model mentioned above, a new extraction network model with better effect can be reconstructed by performing inter-layer scrambling on the 9 convolutional layers in the extraction network model, which is convenient for the implementation of the present invention and safe high degree.

Description of drawings

Fig. 1 is the flow chart of the image steganographic model protection method based on extracting network parameter scrambling in the method embodiment of the present invention;

Fig. 2 is the working process schematic diagram of the image steganography network model in the method embodiment of the present invention;

3 is a schematic diagram of the working effect of an image steganography network model in an embodiment of the method of the present invention;

4 is a schematic diagram of the effect of extracting a single-layer convolution layer scrambling scheme in a network in an embodiment of the method of the present invention;

5 is a schematic diagram of the effect of extracting a multi-layer convolution layer scrambling scheme in a network in an embodiment of the method of the present invention;

FIG. 6 is a schematic diagram of the effect of scrambling the positions of convolution kernels in the last convolution layer in the method embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

Method example:

The present invention performs scrambling operation on the extracted network model through the chaotic system encryption algorithm. Even if the extracted network model is acquired by a third party during the transmission process, the useful secret image cannot be extracted from the encrypted image through the extracted network model. The scrambled extraction network model can be restored to the correct extraction network according to the scrambling algorithm agreed in advance, so as to improve the security and reliability of the communication process between the sender and the receiver.

As shown in FIG. 2 , the present invention constructs an image steganographic network model by using the image steganography model framework of the deep neural convolutional network and the ideas of the encoder and the decoder. The image steganography network model includes a hidden network model (HidingNetwork) and an extraction network model (Extraction Network). The hidden network model uses the encoding function to connect the secret image (Secret image) and the carrier image (Cover image) into a 6-channel tensor, Then a 3-channel container image is output, and the extraction network model can extract the secret information in the container image, so as to recover the secret image (Recovered image) in the container image.

As shown in Figure 1, the image steganographic model protection method based on extracting network parameter scrambling of the present invention comprises the following steps:

1. Build the hidden network model and extract the network model.

The structure of the hidden network model and the extraction network model is shown in Table 1 below. The extraction network model mainly includes three parts, namely the upsampling module, the pooling module, the downsampling module and the convolution module. The upsampling includes 4 The convolutional layers Conv and BN (Batch Normalization) layers are connected to each other, the pooling module includes 1 pooling layer, the downsampling module includes 4 interconnected deconvolution layers DeConv and BN layers, and the convolution module includes 1 layer Interconnected convolutional layers Conv and BN layers. The upsampling module and the downsampling module use the activation function ReLU, and the convolution module uses the activation function Sigmoid.

Table 1. Internal structure table of hidden network model and extracted network model

As can be seen from Table 1, the number of convolution kernels in the first 4 convolutional layers in the extraction network model are 32, 64, 128 and 256, respectively, and the sizes of the convolution kernels are 3×3, 4×4, 4×4 and 4×4; the number of convolution kernels of the last 4 deconvolution layers are 128, 64, 32 and 16, respectively, and the convolution kernel sizes are 4×4, 4×4, 4×4 and 3, respectively ×3; the number of convolution kernels in the last convolutional layer is 3, and the size of the convolution kernel is 3×3. The first four convolution layers perform convolution operations on the input dense image (Container image) in turn, converting the dense image from 3 channels to 256 channel feature maps, and the last four deconvolution layers perform deconvolution operations. The last 1 convolutional layer generates a 3-channel secret image.

100,000 images are randomly selected from the ImageNet dataset, of which 98,000 images are used as the training set, 1,000 are used as the validation set, and the remaining 1,000 are used as the test set. Through 200 epochs training and testing, the steganographic network model and extraction network model of the image steganographic network model are obtained.

As shown in Figure 3, the first row is the carrier image, the second row is the secret image after hiding the secret image, the third row is the secret image that the sender needs to send to the receiver, and the fourth row is the receiver through Extract the secret image extracted by the network model from the secret image. From the content shown in Figure 3, it can be seen that there is almost no gap between the carrier image and the secret image in terms of visual effects, and there is no loss between the extracted secret image and the original hidden secret image, indicating that the trained image steganography The error loss of the network model is smaller, and the hiding effect and extraction effect are better.

2. The sender scrambles the parameters of the trained extraction network model through the agreed scrambling algorithm, and reconstructs the parameters of the scrambled extraction network model to obtain a new extraction network model. Send the encrypted image and the new extracted network model to the recipient.

The specific method is: according to the number of convolution layers and deconvolution layers in the extracted network model, the parameters of the extracted network model are divided into 9 parameter files, and in each parameter file, the convolution kernel is regarded as a minimum operation unit , in MATLAB, the super Lorenz chaotic system scrambling algorithm is used to scramble and encrypt the position of the convolution kernel in each convolution layer and deconvolution layer, so as to obtain a scrambled parameter file, and use the scrambled parameters The file reconstructs a new extraction network model.

Super-Lorenz chaotic systems have distinct nonlinear dynamic characteristics and are sensitive to initial values, so they are widely used in digital image encryption. Usually, the hybrid system scrambles the pixel position of the image to encrypt the image, but this method is the same as file encryption. When the encrypted image is transmitted in an open channel, it is easy to cause suspicion and may be used by a third party. Communication between the two parties is blocked, and the receiver does not receive the sender's information at all.

The system model of the super-Lorenz chaotic system scrambling algorithm is as follows:

In the formula, x, y, z, w are the state variables of the super-Lorenz chaotic system, a, b, c, r are the control parameters of the super-Lorenz chaotic system. When the control parameters of the chaotic system a=10, b=8/3, c=28, -1.52<r≤-0.06, the system is in a hyperchaotic state. A hyperchaotic state is a disordered, unpredictable, and chaotic state.

The state variables x, y, z, and w of the super-Lorenz chaotic system are the keys for scrambled encryption, thereby generating a one-dimensional chaotic sequence. Therefore, the scrambling encryption process is as follows: First, a one-dimensional chaotic sequence is generated through the super-Lorenz chaotic system, and the convolution kernel of each layer is loaded into the super-Lorenz chaotic system, and the convolution kernel in the parameter file is adjusted by the one-dimensional chaotic sequence Scrambling is performed at the position of the scrambled parameter file to obtain a scrambled parameter file, and then a new extraction network model is reconstructed by using the scrambled parameter file. There are two methods for scrambling the position of the convolution kernel:

First, the layer is chaotic. The positions of the convolution kernels in each convolutional layer are scrambled, that is, the positions of the convolution kernels in each convolutional layer and deconvolutional layer are randomly arranged without repetition. For example, when the positions of the 32 convolution kernels in the first convolutional layer are scrambled, and the convolution operation is performed through the first convolutional layer after the dense image is input, the output 32-channel feature map is the same as that of the original first convolutional layer. The 32-channel feature maps output by the stack layer are different. In the same way, if the positions of the convolution kernels in other convolution layers and deconvolution layers are scrambled, the output feature maps will be different when the convolution operation and the deconvolution operation are performed.

As shown in Figure 4, the first row from left to right in the figure is the secret image extracted by the original extraction network model (that is, the extraction network model before scrambling and encryption). The first row is from left to right. The second picture to the fifth picture are the secret images extracted by the extraction network obtained by scrambling the first 4 convolutional layers. The convolutional layer and the last convolutional layer perform layer-in-layer scrambling to obtain the secret image extracted by the extraction network.

It can be seen intuitively from Fig. 4 that the secret images extracted by the extraction network model obtained by the 1st, 2nd, 4th, 7th and 8th convolutional layers by intra-layer scrambling have changed greatly. . To this end, a structural similarity analysis method SSIM (Structural Similarity Index Measurement) is introduced to analyze the similarity between images. The 1st, 2nd, 4th, 7th and 8th convolutional layers perform layer-in-layer scrambling to obtain the secret image extracted from the extraction network model and the original secret image (that is, the secret image extracted by the original extraction network). ), the similarity between the The similarity between the original secret images is higher than the set threshold k. Therefore, if the sender sends the extraction network model obtained by in-layer scrambling of the 1st, 2nd, 4th, 7th or 8th convolutional layer to the receiver, even if the third party obtains the extraction network It is also difficult to obtain a complete secret image from the secret image, and these extraction network models can be used as the extraction network sent to the receiver. In this embodiment, the set threshold value k is set to be 0.1.

Second, inter-layer chaos. Scramble the positions of the convolution kernels in different convolutional layers, that is, randomly select multiple convolutional layers, and randomly arrange the positions of the convolutional kernels in each convolutional layer without repetition. The position of the kernel changes not only within the layer, but also to other convolutional or deconvolutional layers.

In this embodiment, 4 secret images extracted by the extraction network model after inter-layer scrambling are provided for display, respectively: 1) Select the 1st, 2nd, 4th, 7th and 8th images The convolution kernel of the convolutional layer performs inter-layer scrambling; 2) selects the first 4 convolutional layers for inter-layer scrambling; 3) selects the last 4 deconvolution layers for inter-layer scrambling; 4) selects all the The convolutional and deconvolutional layers perform inter-layer scrambling.

As shown in Figure 5, the first image is the original secret image, and the second to fifth images are the secret images extracted from the secret image by the extraction network model obtained after the above four inter-layer scrambling. Similarly, the structural similarity analysis method SSIM is used to analyze the similarity, and it can be concluded that the similarity between each secret image and the original secret image is less than the set threshold k. Therefore, if the sender sends the extracted network model obtained after the above four inter-layer scrambling to the receiver, even if a third party obtains the extracted network model, it is difficult to obtain a complete secret image from the encrypted image. The network model can then be used as the extracted network model sent to the receiver.

In addition, as can be seen from the last image in Figure 4, when the last convolutional layer is scrambled, the resulting extraction network model may only affect the color of the image when extracting the secret image. For this reason, the last convolutional layer is scrambled separately. The last convolutional layer has only 3 convolution kernels, its initial position is 123, and the position of the convolution kernel after the scramble in the layer is only 5 cases: 132 , 213, 231, 312 and 321. As shown in Figure 6, scrambling the position of the convolution kernel in the last convolutional layer alone will only make the secret image extracted by the extraction network model different from the original secret image, and will still make the information in the secret image different. Expose, therefore, does not use the extraction network model obtained by in-layer scrambling of the positions of the convolution kernels in the last convolutional layer alone.

The sender sends the control parameters a=10, b=8/3, c=28, -1.52<r≤-0.06 and state variables of the hyperchaotic scrambling system to the receiver, and agrees with the receiver on the scrambling algorithm. The method is intra-layer scrambling or inter-layer scrambling, and which convolutional or deconvolutional layers are scrambled, so as to realize the advance agreement of the scrambling algorithm.

As other implementation manners, a relatively simple scrambling method may also be agreed upon. In this case, the sender and receiver do not need to agree on control parameters in advance. For example, for intra-layer scrambling, the positions of the convolution kernels in a convolutional layer are exchanged according to the agreed rules, such as the end-to-end exchange or the symmetrical exchange about the middle position of the convolution kernels; for inter-layer scrambling, some two layers The position of the convolution kernel at the same position in the convolutional layer is swapped.

In order to improve the security of the subsequent image steganography process, the sender can also regularly re-agree with the receiver on the scrambling algorithm.

3. After receiving the extracted network model, the receiver first reconstructs the chaotic scrambling system according to the control parameters a=10, b=8/3, c=28, -1.52<r≤-0.06 of the chaotic scrambling system sent by the sender in advance. Then use the received state variable as the key and the received extraction network as the object to be decrypted, reconstruct the chaotic scrambled system model from the state variable and the extracted network input, and use the agreed scrambling algorithm Scramble and decrypt the extracted network model and restore it to the original extracted network model. Therefore, the receiver can use the original extraction network model to extract the secret image with the secret image hidden, so as to obtain the secret image.

By adopting the invention, the extraction network model in the image steganography network model can be encrypted, and the security of image steganography is further improved.

Claims (10)

1. An image steganography model protection method based on extraction network parameter scrambling is characterized by comprising the following steps:

s1, hiding the secret image to be sent into a carrier image by a sender through the trained hidden network model to obtain a corresponding secret-carrying image, and obtaining a trained extraction network model, wherein the hidden network model and the extraction network model jointly form an image steganography network for deep learning;

s2, the sender scrambles the trained parameters of the extraction network model through an agreed scrambling algorithm, and reconstructs the parameters according to the scrambled parameters of the extraction network model to obtain a new extraction network model; the sender sends the secret-carrying image and the new extraction network model to the receiver; the promised scrambling algorithm satisfies the following conditions: for the same secret-carrying image, the similarity of the secret image extracted by the extraction network model before scrambling and the secret image extracted by the extraction network model after scrambling is lower than a set threshold value;

and S3, after receiving the new extracted network model, the receiver scrambles and decrypts the new extracted network model according to the agreed scrambling algorithm to obtain the extracted network model before scrambling, and extracts the received secret-carrying image according to the extracted network model before scrambling to obtain the secret image.

2. The image steganography model protection method based on extraction network parameter scrambling of claim 1, wherein the steganography network model and the extraction network model both adopt a deep neural convolution network, and the scrambling operation refers to performing in-layer scrambling on positions of convolution kernels in one or more convolutional layers in the trained extraction network model or performing interlayer scrambling on positions of convolution kernels in different convolutional layers; the position of the convolution kernel is changed only in the convolution layer when the layers are scrambled, and the position of the convolution kernel is changed in different convolution layers when the layers are scrambled.

3. The image steganography model protection method based on extraction network parameter scrambling of claim 2, wherein the agreed scrambling algorithm employs a hyper-Lorenz chaotic system.

4. The image steganography model protection method based on extraction network parameter scrambling as claimed in claim 1, wherein the similarity is determined by using structural similarity analysis method SSIM.

5. The image steganography model protection method based on extraction network parameter scrambling of claim 2, wherein the extraction network model comprises an upsampling module, a pooling layer, a downsampling module, and a convolution layer connected to each other; the upsampling module and the downsampling module each include 4 convolutional layers.

6. The image steganography model protection method based on extraction network parameter scrambling of claim 5, wherein in the scrambling operation in step S2, the 1 st, 2 nd or 4 th convolution layer of the up-sampling module is subjected to intra-layer scrambling, or the 3 rd or 4 th convolution layer of the down-sampling module is subjected to intra-layer scrambling.

7. The image steganography model protection method based on extraction network parameter scrambling of claim 5, wherein in the scrambling operation of step S2, the 1 st, 2 nd and 4 th convolution layers of the up-sampling module and the 3 rd and 4 th convolution layers of the down-sampling module are subjected to interlayer scrambling.

8. The image steganography model protection method based on extraction network parameter scrambling of claim 5, wherein in the scrambling operation in step S2, the 4 convolution layers of the upsampling module are subjected to interlayer scrambling.

9. The image steganography model protection method based on extracted network parameter scrambling as claimed in claim 5, wherein in step S2, when the scrambling operation is performed, the inter-layer scrambling is performed on 4 convolution layers of the down-sampling module.

10. The method for protecting an image steganography model based on extracted network parameter scrambling of claim 5, wherein in the scrambling operation in step S2, the layers of 9 convolution layers in the extracted network model are scrambled.

Images