WO2022237078A1

WO2022237078A1 - Color image steganography method based on generative adversarial network

Info

Publication number: WO2022237078A1
Application number: PCT/CN2021/125637
Authority: WO
Inventors: 李京兵; 曾城; 刘婧; 黄梦醒
Original assignee: 海南大学
Priority date: 2021-05-13
Filing date: 2021-10-22
Publication date: 2022-11-17
Also published as: CN113077377B; CN113077377A

Abstract

A color image steganography method based on a generative adversarial network. The method comprises: manufacturing a carrier image data set and a secret image data set, and dividing each of the carrier image data set and the secret image data set into a training set, a verification set and a test set (S101); constructing an image steganography model, which is composed of a pre-processing network, a hiding network, an extraction network and a steganalysis network (S102), wherein the pre-processing network is used for pre-processing a carrier image and a secret image, the hiding network is used for hiding the secret image in the carrier image, so as to obtain a steganographic image, the extraction network is used for obtaining an extracted image from the steganographic image, and the steganalysis network is used for performing steganalysis on the carrier image and the steganographic image; performing training and parameter tuning on the model by using the training set and the verification set, and forming adversarial training by using the hiding network and the steganalysis network (S103); and performing a performance test on the model by using the test set (S104). In this way, a high-quality steganographic image and extracted image can be obtained, and the steganographic image has an anti-steganalysis capability.

Description

A Color Image Steganography Method Based on Generative Adversarial Networks

This application claims the priority of the Chinese patent application with the application number 202110523597.8 and the invention title "A Color Image Steganography Method Based on Generative Adversarial Network" submitted to the China Patent Office on May 13, 2021, the entire contents of which are incorporated by reference incorporated in this application.

technical field

The invention relates to the technical field of signal processing, in particular to a color image steganography method based on generating an adversarial network.

Background technique

With the development of computer technology, the transmission of information is more and more convenient, but at the same time, the security of information is also facing great challenges. To deal with these challenges, researchers have proposed many measures, such as digital watermarking, cryptography and information hiding. Information hiding has always been an important field of information security research. Compared with other fields, such as cryptography is to prevent secret messages from being discovered, while information hiding needs to hide the process of information transmission. Compared with text and audio, images have rich redundant information, and are usually used as the carrier of information hiding. At present, most information hiding research uses images as carriers, while secret messages can be images, texts, or audios. Image steganography can be seen as a disguised encryption technique, but it is different from traditional encryption techniques. Image steganography embeds messages into images without changing their perceptual characteristics, ensuring that the carrier embedded with secret information looks visually It is very similar to the original image, so as to avoid the attention and suspicion of the monitor, and avoid being intercepted and attacked by the monitor.

Convolutional neural network has superior feature extraction and feature representation capabilities, and has been widely used in tasks such as computer vision and natural language processing, and has achieved remarkable results. Therefore, many scholars try to use the powerful feature extraction ability of deep learning to apply deep learning theory to image steganography. The evaluation indicators of image steganography mainly include imperceptibility, security and steganographic capacity. The imperceptibility mainly refers to the fact that the steganographic image cannot be detected by human eyes, the security mainly refers to the ability to resist certain anti-steganographic analysis, and the steganographic capacity refers to the ability to embed enough information in the carrier. However, in traditional image steganography methods, the more secret information is embedded, the less imperceptible the image will be.

Therefore, how to balance the security of the image and the steganography capacity is a technical problem to be solved urgently by those skilled in the art.

Contents of the invention

Based on this, it is necessary to address the above technical problems and provide a color image steganography method based on generative adversarial networks, which can obtain high-quality steganographic images and extracted images, and steganographic images have a certain ability to resist steganalysis. The specific plan is as follows:

A color image steganography method based on generative confrontation network, including:

Making a carrier image data set and a secret image data set, dividing the carrier image data set and the secret image data set into a training set, a verification set and a test set respectively; the carrier image in the carrier image data set and the secret image data The secret images in the set are of the same size;

Construct an image steganographic model consisting of a preprocessing network, a hidden network, an extraction network and a steganalysis network; the preprocessing network is used to preprocess the carrier image and the secret image; the hidden network is used to hiding the secret image in the carrier image to obtain a steganographic image; the extraction network is used to obtain an extracted image from the steganographic image; the steganalysis network is used to analyze the carrier image and the steganographic image Steganalysis of steganographic images;

The image steganography model constructed is trained and parameter-tuned through the training set and the verification set; the hidden network and the steganalysis network form an adversarial training;

Using the test set to test the performance of the image steganography model after parameter tuning.

Preferably, in the above-mentioned color image steganography method based on generative confrontation network provided by the embodiment of the present invention, the training of the constructed image steganography model includes:

Inputting the carrier image and the secret image into the preprocessing network respectively to obtain a first feature map;

inputting the first feature map into the hidden network to obtain the steganographic image;

inputting the steganographic image into the extraction network to obtain the extracted image;

Inputting the carrier image and the steganographic image as input images into the steganalysis network to obtain an analysis result of the input image.

Preferably, in the above-mentioned color image steganography method based on generative adversarial networks provided by the embodiments of the present invention, the carrier image and the secret image are respectively input into the preprocessing network to obtain a first feature map, include:

The cover image and the secret image are respectively input into the two branches of the preprocessing network, and the cover image and the secret image pass through the first convolutional layer of each branch respectively, and are activated with a ReLU activation function , get the first carrier feature map and the first secret feature map respectively;

passing the first carrier feature map and the first secret feature map through the second convolutional layer of each branch, and activating with a ReLU activation function to obtain a second carrier feature map and a second secret feature map;

passing the second carrier feature map and the second secret feature map through the third convolutional layer of each branch, and activating with a ReLU activation function to obtain a third carrier feature map and a third secret feature map;

Combining the third carrier feature map and the third secret feature map to obtain a first feature map.

Preferably, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the input of the first feature map into the hidden network to obtain the steganographic image includes:

The first feature map is input to the first convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain a second feature map;

The first feature map is merged with the second feature map by skip connection, and input to the second convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain the third feature picture;

The first feature map and the second feature map are merged by skip connection and the third feature map respectively, and input to the third convolutional layer of the hidden network for convolution and using ReLU The function is activated to obtain the fourth feature map;

Combining the first feature map, the second feature map and the third feature map through skip connections and the fourth feature map respectively, and inputting them into the fourth convolutional layer of the hidden network Perform convolution and use the ReLU function for activation to obtain the fifth feature map;

Merging the first feature map, the second feature map, the third feature map and the fourth feature map through skip connections and the fifth feature map respectively, and inputting them into the hidden network In the fifth convolutional layer of the convolution and activation using the ReLU function, the sixth feature map is obtained;

merging the first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map via skip connections and the sixth feature map, respectively, And input to the sixth convolutional layer of the hidden network for convolution and activation using the Tanh function to obtain the seventh feature map;

The steganographic image is obtained by adding the cover image to the seventh feature map through skip connections.

Preferably, in the above-mentioned color image steganography method based on GAN provided by an embodiment of the present invention, the inputting the steganographic image into the extraction network to obtain the extraction image includes:

The steganographic image is input to the first convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the first steganographic feature map;

Inputting the first steganographic feature map into the second convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a second steganographic feature map;

Inputting the second steganographic feature map into the third convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a third steganographic feature map;

Inputting the third steganographic feature map into the fourth convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a fourth steganographic feature map;

Inputting the fourth steganographic feature map into the fifth convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a fifth steganographic feature map;

The fifth steganographic feature map is input to the sixth convolutional layer of the extraction network for convolution and activation using a Tanh function to obtain the extracted image.

Preferably, in the above-mentioned color image steganography method based on generative adversarial networks provided by the embodiments of the present invention, the carrier image and the steganographic image are input into the steganalysis network as input images to obtain the The analysis results of the above input image, including:

Inputting the input image into the first convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a first input feature map;

Inputting the first input feature map into the second convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a second input feature map;

Inputting the second input feature map into the third convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain a third input feature map;

Inputting the third input feature map into the fourth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a fourth input feature map;

Inputting the fourth input feature map into the fifth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a fifth input feature map;

Inputting the fifth input feature map into the sixth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a sixth input feature map;

The sixth input feature map is input to the pyramid pooling layer to obtain the first vector;

The first vector is input to the fully connected layer, activated by a Sigmoid function, and the second vector is obtained as an analysis result of the input image.

Preferably, in the above-mentioned color image steganography method based on generative confrontation network provided by the embodiment of the present invention, the convolution kernels of the preprocessing network are all 3x3, and the step size is 1; each branch of the preprocessing network The number of convolution kernels in the first convolution layer is 128, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer is 32;

The convolution kernels of the hidden network are all 3x3, and the step size is 1; the number of convolution kernels in the first convolution layer of the hidden network is 32, and the number of convolution kernels in the second convolution layer The number of convolution kernels in the third convolution layer is 128, the number of convolution kernels in the fourth convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 64. The number of convolution kernels in each convolutional layer is 3;

The convolution kernels of the extraction network are all 3x3, and the step size is 1; the number of convolution kernels in the first convolution layer of the extraction network is 32, and the number of convolution kernels in the second convolution layer The number of convolution kernels in the third convolution layer is 128, the number of convolution kernels in the fourth convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 64. The number of convolution kernels in each convolutional layer is 3;

The convolution kernels of the steganalysis network are all 1x1, and the step size is 2; the number of convolution kernels in the first convolution layer of the steganalysis network is 24, and the number of convolution kernels in the second convolution layer is 24. The number of convolution kernels is 48, the number of convolution kernels in the third convolution layer is 96, the number of convolution kernels in the fourth convolution layer is 128, and the number of convolution kernels in the fifth convolution layer is 128. The number of convolution kernels in the sixth convolution layer is 128.

Preferably, in the above-mentioned GAN-based color image steganography method provided in an embodiment of the present invention, the parameter tuning of the constructed image steganography model includes:

Tuning the learning rate of the constructed image steganography network model, the number of images to be processed in batches, the iteration cycle and the optimizer.

Preferably, in the above-mentioned GAN-based color image steganography method provided in an embodiment of the present invention, the performance test of the image steganography model after parameter tuning includes:

Input the carrier image and the secret image in the test set into the image steganography network model after parameter tuning respectively, and output the steganographic image and the extracted image corresponding to the test set;

Carrying out a quality test on the steganographic image and the extracted image corresponding to the output test set;

Performing a security test on the steganographic image corresponding to the output test set;

The generalization ability test is carried out on the image steganography model after parameter tuning.

Preferably, in the above-mentioned GAN-based color image steganography method provided in an embodiment of the present invention, the quality test is performed on the output steganographic image and the extracted image corresponding to the test set, including:

Calculate the output PSNR value and SSIM value between the steganographic image corresponding to the test set and the cover image in the test set, and calculate and output the extracted image corresponding to the test set and the secret image in the test set PSNR value and SSIM value between;

According to the calculated PSNR value and SSIM value, the quality test result is obtained.

It can be seen from the above technical solutions that a color image steganography method based on generative adversarial networks provided by the present invention includes: making a carrier image dataset and a secret image dataset, and separating the carrier image dataset and the secret image dataset Divided into training set, validation set and test set; the cover image in the cover image dataset and the secret image in the secret image dataset have the same size; build an image steganography network consisting of a preprocessing network, a hidden network, an extraction network and a steganalysis network model; the preprocessing network is used to preprocess the cover image and the secret image; the hidden network is used to hide the secret image into the cover image to obtain the steganographic image; the extraction network is used to extract the image from the steganographic image; the steganographic The analysis network is used to conduct steganalysis on the carrier image and the steganographic image; the image steganography model constructed through the training set and the verification set is trained and the parameters are tuned; the hidden network and the steganalysis network form an adversarial training; the test set is used The performance test of the image steganography model after parameter tuning is carried out.

The image steganography model finally obtained by the above-mentioned color image steganography method provided by the present invention can embed color images into color images of the same size, and can obtain high-quality steganographic images and extracted images through confrontation training, and simultaneously steganographic images It has a certain ability to resist steganalysis.

The above description is only an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the description, and in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable , the specific embodiments of the present invention are enumerated below.

Description of drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiment. The drawings are only for the purpose of illustrating a preferred embodiment and are not to be considered as limiting the invention. Also throughout the drawings, the same reference numerals are used to designate the same parts. In the attached picture:

FIG. 1 shows a flow chart of a color image steganography method based on a generative adversarial network provided by an embodiment of the present invention;

FIG. 2 shows a schematic structural diagram of an image steganography model provided by an embodiment of the present invention;

FIG. 3 shows a schematic diagram of a specific structure of an image steganography model provided by an embodiment of the present invention;

Fig. 4 shows the anti-steganalysis ROC diagram provided by the embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The present invention provides a kind of color image steganography method based on generation confrontation network, as shown in Figure 1, comprises the following steps:

S101. Make a carrier image data set and a secret image data set, divide the carrier image data set and the secret image data set into a training set, a verification set and a test set respectively; the carrier image in the carrier image data set and the secret image in the secret image data set same size.

Specifically, 24,000 images can be randomly selected from the ImageNet dataset, and the size of the images is adjusted to the same size (such as 256x256), and divided into a cover image dataset and a secret image dataset; and then these two datasets are divided into three disjoint parts, where the training set has a total of 20,000 images, the validation set has a total of 2,000 images, and the test set has a total of 2,000 images.

S102. Construct an image steganography model consisting of a preprocessing network, a hidden network, an extraction network, and a steganalysis network; the preprocessing network is used to preprocess the carrier image and the secret image to improve the hiding efficiency of the hidden network; the hidden network is used to It is used to hide the secret image into the cover image to obtain the steganographic image; the extraction network is used to obtain the extracted image from the steganographic image; the steganalysis network is used to steganalyze the cover image and the steganographic image.

It should be noted that the preprocessing network has two input branches and one output; the input of the hidden network is the output of the preprocessing network, and the output is the steganographic image; the extraction network extracts the secret image from the steganographic image; the steganalysis network can Determine whether the input image is a natural image or a steganographic image.

S103. Perform training and parameter tuning on the constructed image steganography model through the training set and the verification set; the hidden network and the steganalysis network form an adversarial training.

It should be noted that the overall parameter tuning of the model should make the model loss function as small as possible. The hidden network uses a dense residual module, which can transfer the low-level features of the image to each subsequent layer, which helps to hide the details of the secret image. Simultaneous adversarial training of hidden network and steganalysis network can improve the quality and security of steganographic images.

S104. Use the test set to perform a performance test on the image steganography model after parameter tuning.

In the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the finally obtained image steganography model can embed color images into color images of the same size, and high-quality steganography can be obtained through adversarial training Image and image extraction, while the steganographic image has a certain ability to resist steganalysis.

In specific implementation, in the above-mentioned color image steganography method based on generative confrontation network provided by the embodiment of the present invention, as shown in FIG. 2 , step S103 trains the constructed image steganography model, which may include the following steps:

Step 1. Input the carrier image C(0) and the secret image S(0) respectively into the preprocessing network to obtain the first feature map I(1);

Step 2. Input the first feature map I(1) into the hidden network to obtain the steganographic image C'(0);

Step 3: Input the steganographic image C'(0) into the extraction network to obtain the extracted image S'(0);

Step 4: Input the carrier image C(0) and the steganographic image C'(0) as the input image R(0) into the steganalysis network to obtain the analysis result of the input image R(0).

In specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, as shown in FIG. To the preprocessing network, the first feature map I(1) is obtained, which may specifically include the following steps:

The cover image C(0) and the secret image S(0) are respectively input into the two branches of the preprocessing network, and the cover image C(0) and the secret image S(0) respectively pass through the first convolutional layer of each branch , activate with the ReLU activation function to obtain the first carrier feature map C(1) and the first secret feature map S(1);

Pass the first carrier feature map C(1) and the first secret feature map S(1) respectively through the second convolutional layer of each branch, and use the ReLU activation function to activate the second carrier feature map C(2) and The second secret feature map S(2);

Pass the second carrier feature map C(2) and the second secret feature map S(2) respectively through the third convolutional layer of each branch, and use the ReLU activation function to activate, respectively, to obtain the third carrier feature map C(3) and The third secret feature map S(3);

Merge the third carrier feature map C(3) and the third secret feature map S(3) to obtain the first feature map I(1).

Further, in specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the convolution kernels of the preprocessing network are all 3x3, and the step size is 1; as shown in Figure 3 , the number of convolution kernels in the first convolution layer of each branch of the preprocessing network is 128, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer for 32.

In specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, as shown in FIG. Writing the image C'(0) may specifically include the following steps:

Input the first feature map I(1) into the first convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain the second feature map I(2);

The first feature map I(1) is merged with the second feature map I(2) by skip connection, and input to the second convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain the first Three feature map I(3);

The first feature map I(1) and the second feature map I(2) are merged by skip connection and the third feature map I(3), respectively, and input to the third convolutional layer of the hidden network for convolution The product sum is activated using the ReLU function to obtain the fourth feature map I(4);

Merge the first feature map I(1), the second feature map I(2) and the third feature map I(3) through skip connections and the fourth feature map I(4) respectively, and input them into the hidden network Convolution is performed in the fourth convolutional layer and activation is performed using the ReLU function to obtain the fifth feature map I(5);

The first feature map I(1), the second feature map I(2), the third feature map I(3) and the fourth feature map I(4) are respectively passed through the skip connection and the fifth feature map I(5) Merge and input to the fifth convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain the sixth feature map I(6);

The first feature map I(1), the second feature map I(2), the third feature map I(3), the fourth feature map I(4) and the fifth feature map I(5) are respectively connected by skipping Merge with the sixth feature map I(6), and input it to the sixth convolutional layer of the hidden network for convolution and activation using the Tanh function to obtain the seventh feature map I(7);

Add the cover image C(0) to the seventh feature map I(7) through skip connection to obtain the steganographic image C'(0).

Further, in specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the convolution kernels of the hidden network are all 3x3, and the step size is 1; as shown in Figure 3, The number of convolution kernels in the first convolution layer of the hidden network is 32, the number of convolution kernels in the second convolution layer is 64, the number of convolution kernels in the third convolution layer is 128, and the number of convolution kernels in the fourth convolution layer is The number of convolution kernels in the first convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 3.

In specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, as shown in Figure 3, in step 3, the steganographic image C'(0) is input to the extraction network to obtain the extraction The image S'(0) may specifically include the following steps:

Input the steganographic image C'(0) into the first convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the first steganographic feature map D(1);

Input the first steganographic feature map D(1) into the second convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the second steganographic feature map D(2);

Input the second steganographic feature map D(2) into the third convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the third steganographic feature map D(3);

Input the third steganographic feature map D(3) into the fourth convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the fourth steganographic feature map D(4);

Input the fourth steganographic feature map D(4) into the fifth convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the fifth steganographic feature map D(5);

Input the fifth steganographic feature map D(5) into the sixth convolutional layer of the extraction network for convolution and activation using the Tanh function to obtain the extracted image S'(0).

Further, in specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the convolution kernels of the extraction network are all 3x3, and the step size is 1; as shown in Figure 3, The number of convolution kernels in the first convolution layer of the extraction network is 32, the number of convolution kernels in the second convolution layer is 64, the number of convolution kernels in the third convolution layer is 128, and the number of convolution kernels in the fourth convolution layer is The number of convolution kernels in the first convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 3.

In specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, as shown in Figure 3, in step 4, the carrier image C(0) and the steganographic image C'(0) As the input image R(0) is input to the steganalysis network to obtain the analysis result of the input image R(0), which may specifically include the following steps:

Input the input image R(0) into the first convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the first input feature map R(1);

Input the first input feature map R(1) into the second convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the second input feature map R(2);

Input the second input feature map R(2) into the third convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the third input feature map R(3);

Input the third input feature map R(3) into the fourth convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the fourth input feature map R(4);

Input the fourth input feature map R(4) into the fifth convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the fifth input feature map R(5);

Input the fifth input feature map R(5) into the sixth convolutional layer of the steganalysis network for convolution and activation using the ReLU function to obtain the sixth input feature map R(6);

Input the sixth input feature map R(6) to the pyramid pooling layer to obtain the first vector R(7);

Input the first vector R(7) into the fully connected layer, use the Sigmoid function to activate, and obtain the second vector R(8) as the analysis result of the input image R(0), that is, the evaluation score of the input image R(0) .

Further, in specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, the convolution kernel of the steganalysis network is 1x1, and the step size is 2; as shown in Figure 3 It shows that the number of convolution kernels in the first convolution layer of the steganalysis network is 24, the number of convolution kernels in the second convolution layer is 48, and the number of convolution kernels in the third convolution layer is 96. The number of convolution kernels in the fourth convolution layer is 128, the number of convolution kernels in the fifth convolution layer is 128, and the number of convolution kernels in the sixth convolution layer is 128.

In specific implementation, in the above-mentioned color image steganography method based on generative confrontation network provided by the embodiment of the present invention, in the process of constructing the image steganography model, it is necessary to set the relevant parameters of model training, that is, to set the learning rate of the network model, Set the batch_size of images processed in a batch in the network model, set the iteration cycle epoch, set the optimizer of the network model optimizer, and set the training steps of the network, which can be trained in two steps: the first step is to train the steganalysis network, The second step trains the preprocessing network, hidden network and extraction network. Next, in the process of tuning the parameters of the constructed image steganography model, it may include: the learning rate of the constructed image steganography network model, the number of batches of images batch_size, the iteration cycle epoch and the optimizer optimizer These parameters are tuned.

During specific implementation, in the above-mentioned color image steganography method based on generative adversarial network provided by the embodiment of the present invention, step S104 performs a performance test on the image steganography model after parameter tuning, which may specifically include: first, centralize the test The cover image and the secret image of the parameter are respectively input into the image steganographic network model after parameter tuning, and the steganographic image and the extracted image corresponding to the test set are output; then, the quality of the steganographic image and the extracted image corresponding to the output test set are Test; conduct a security test on the steganographic image corresponding to the output test set; conduct a generalization ability test on the image steganography model after parameter tuning. The resulting steganographic image has high quality and security, and the model has good generalization ability.

Further, in specific implementation, in the above-mentioned GAN-based color image steganography method provided by the embodiment of the present invention, in the above steps, the quality test of the steganographic image and the extracted image corresponding to the output test set may include : Calculate the PSNR (peak signal-to-noise ratio) value and SSIM (structural similarity) value between the steganographic image corresponding to the output test set and the cover image in the test set, and calculate the extracted image and test corresponding to the output test set The PSNR value and SSIM value between the concentrated secret images; according to the calculated PSNR value and SSIM value, the quality test result is obtained.

Specifically, the test set of the ImageNet dataset is tested, assuming that the test set has a total of 2000 images, of which 1000 images can be selected as carrier images, and the other 1000 images are secret images. After testing, the quality test results are: the PSNR and SSIM of the carrier image and the steganographic image are 44.15 and 0.99, respectively, and the PSNR and SSIM of the secret image and the extracted image are 40.03 and 0.98, respectively, indicating that the image steganography model after parameter tuning High-quality steganographic images and extracted images can be obtained.

In the process of performing the security test on the steganographic image corresponding to the output test set in the above steps, the steganalysis tool StegExpose can be used to perform steganographic analysis on the image, and the result is shown in Figure 4, a straight line in Figure 4 Indicates random prediction, and the curve on the upper left of the straight line represents the steganalysis of the image. It can be found that the curve is close to the straight line, indicating that StegExpose's steganalysis of the steganographic image is close to random prediction, indicating that the steganographic image has a certain anti-steganographic analysis ability.

In the process of testing the generalization ability of the image steganography model after parameter tuning in the above steps, 2000 images can be selected from each of the three data sets of LFW, VOC2007 and CoCo to verify the generalization of the image steganography model capabilities, as shown in Table 1:

Table I

It can be seen from Table 1 that the image steganography model has good generalization ability.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other.

Professionals can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the possible For interchangeability, in the above description, the composition and steps of each example have been generally described according to their functions. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present application.

The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be directly implemented by hardware, software modules executed by a processor, or a combination of both. Software modules can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other Any other known storage medium.

A color image steganography method based on generative adversarial networks provided by an embodiment of the present invention includes: making a carrier image data set and a secret image data set, and dividing the carrier image data set and the secret image data set into a training set and a verification set respectively and the test set; the size of the cover image in the cover image dataset and the secret image in the secret image dataset are the same; construct an image steganographic model consisting of a preprocessing network, a hidden network, an extraction network and a steganalysis network; the preprocessing network is used for Preprocess the cover image and the secret image; the hidden network is used to hide the secret image into the cover image to obtain the steganographic image; the extraction network is used to extract the image from the steganographic image; the steganalysis network is used to analyze the cover image Steganalysis with steganographic images; training and parameter tuning of the constructed image steganographic model through the training set and verification set; hidden network and steganalysis network form confrontation training; use the test set to optimize the parameters of the image Steganography model for performance testing. The image steganography model finally obtained through the above-mentioned color image steganography method provided by the present invention can embed color images into color images of the same size, and can obtain high-quality steganographic images and extracted images through confrontational training, and simultaneously steganographic images It has a certain ability to resist steganalysis.

Finally, it should also be noted that in this text, relational terms such as first and second etc. are only used to distinguish one entity or operation from another, and do not necessarily require or imply that these entities or operations, any such actual relationship or order exists. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still carry out the foregoing embodiments Modifications to the technical solutions recorded in the examples, or equivalent replacement of some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should be included in within the protection scope of the present invention.

Claims

A color image steganography method based on generative confrontation network, characterized in that it includes:

Making a carrier image data set and a secret image data set, dividing the carrier image data set and the secret image data set into a training set, a verification set and a test set respectively; the carrier image in the carrier image data set and the secret image data The secret images in the set are of the same size;

Construct an image steganographic model consisting of a preprocessing network, a hidden network, an extraction network and a steganalysis network; the preprocessing network is used to preprocess the carrier image and the secret image; the hidden network is used to hiding the secret image in the carrier image to obtain a steganographic image; the extraction network is used to obtain an extracted image from the steganographic image; the steganalysis network is used to analyze the carrier image and the steganographic image Steganalysis of steganographic images;

The image steganography model constructed is trained and parameter-tuned through the training set and the verification set; the hidden network and the steganalysis network form an adversarial training;

Using the test set to test the performance of the image steganography model after parameter tuning.
The color image steganography method based on generation confrontation network according to claim 1, wherein said training the image steganography model constructed comprises:

Inputting the carrier image and the secret image into the preprocessing network respectively to obtain a first feature map;

inputting the first feature map into the hidden network to obtain the steganographic image;

inputting the steganographic image into the extraction network to obtain the extracted image;

Inputting the carrier image and the steganographic image as input images into the steganalysis network to obtain an analysis result of the input image.
The color image steganography method based on generating an adversarial network according to claim 2, wherein said inputting said carrier image and said secret image into said preprocessing network respectively to obtain a first feature map, comprising :

The cover image and the secret image are respectively input into the two branches of the preprocessing network, and the cover image and the secret image pass through the first convolutional layer of each branch respectively, and are activated with a ReLU activation function , get the first carrier feature map and the first secret feature map respectively;

The first carrier feature map and the first secret feature map are respectively passed through the second convolution layer of each branch, activated with a ReLU activation function, and the second carrier feature map and the second secret feature map are respectively obtained;

passing the second carrier feature map and the second secret feature map through the third convolutional layer of each branch, and activating with a ReLU activation function to obtain a third carrier feature map and a third secret feature map;

Combining the third carrier feature map and the third secret feature map to obtain a first feature map.
The color image steganography method based on generating an adversarial network according to claim 3, wherein said inputting said first feature map into said hidden network to obtain said steganographic image comprises:

The first feature map is input to the first convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain a second feature map;

The first feature map is merged with the second feature map by skip connection, and input to the second convolutional layer of the hidden network for convolution and activation using the ReLU function to obtain the third feature picture;

The first feature map and the second feature map are merged by skip connection and the third feature map respectively, and input to the third convolutional layer of the hidden network for convolution and using ReLU The function is activated to obtain the fourth feature map;

Combining the first feature map, the second feature map and the third feature map through skip connections and the fourth feature map respectively, and inputting them into the fourth convolutional layer of the hidden network Perform convolution and use the ReLU function for activation to obtain the fifth feature map;

Merging the first feature map, the second feature map, the third feature map and the fourth feature map through skip connections and the fifth feature map respectively, and inputting them into the hidden network In the fifth convolutional layer of the convolution and activation using the ReLU function, the sixth feature map is obtained;

merging the first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map via skip connections and the sixth feature map, respectively, And input to the sixth convolutional layer of the hidden network for convolution and activation using the Tanh function to obtain the seventh feature map;

Adding the carrier image to the seventh feature map through skip connections to obtain the steganographic image.
The color image steganography method based on generating an adversarial network according to claim 4, wherein said inputting said steganographic image into said extraction network to obtain said extraction image comprises:

The steganographic image is input into the first convolutional layer of the extraction network for convolution and activation using the ReLU function to obtain the first steganographic feature map;

Inputting the first steganographic feature map into the second convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a second steganographic feature map;

Inputting the second steganographic feature map into the third convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a third steganographic feature map;

Inputting the third steganographic feature map into the fourth convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a fourth steganographic feature map;

Inputting the fourth steganographic feature map into the fifth convolutional layer of the extraction network for convolution and activation using a ReLU function to obtain a fifth steganographic feature map;

The fifth steganographic feature map is input to the sixth convolutional layer of the extraction network for convolution and activation using a Tanh function to obtain the extracted image.
The color image steganography method based on generative adversarial networks according to claim 5, wherein the steganalysis network is input with the carrier image and the steganographic image as input images to obtain the The analysis results of the input image, including:

Inputting the input image into the first convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a first input feature map;

Inputting the first input feature map into the second convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a second input feature map;

Inputting the second input feature map into the third convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a third input feature map;

Inputting the third input feature map into the fourth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a fourth input feature map;

Inputting the fourth input feature map into the fifth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a fifth input feature map;

Inputting the fifth input feature map into the sixth convolutional layer of the steganalysis network for convolution and activation using a ReLU function to obtain a sixth input feature map;

The sixth input feature map is input to the pyramid pooling layer to obtain the first vector;

The first vector is input to the fully connected layer, activated by a Sigmoid function, and the second vector is obtained as an analysis result of the input image.
The color image steganography method based on generating an adversarial network according to claim 6, wherein the convolution kernels of the preprocessing network are all 3x3, and the step size is 1; each branch of the preprocessing network The number of convolution kernels in the first convolution layer is 128, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer is 32;

The convolution kernels of the hidden network are all 3x3, and the step size is 1; the number of convolution kernels in the first convolution layer of the hidden network is 32, and the number of convolution kernels in the second convolution layer The number of convolution kernels in the third convolution layer is 128, the number of convolution kernels in the fourth convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 64. The number of convolution kernels in each convolutional layer is 3;

The convolution kernels of the extraction network are all 3x3, and the step size is 1; the number of convolution kernels in the first convolution layer of the extraction network is 32, and the number of convolution kernels in the second convolution layer The number of convolution kernels in the third convolution layer is 128, the number of convolution kernels in the fourth convolution layer is 64, the number of convolution kernels in the fifth convolution layer is 32, and the number of convolution kernels in the sixth convolution layer is 64. The number of convolution kernels in each convolutional layer is 3;

The convolution kernels of the steganalysis network are all 1x1, and the step size is 2; the number of convolution kernels in the first convolution layer of the steganalysis network is 24, and the number of convolution kernels in the second convolution layer is 24. The number of convolution kernels is 48, the number of convolution kernels in the third convolution layer is 96, the number of convolution kernels in the fourth convolution layer is 128, and the number of convolution kernels in the fifth convolution layer is 128. The number of convolution kernels in the sixth convolution layer is 128.
The method for color image steganography based on generative confrontation network according to claim 1, wherein said parameter tuning of the constructed image steganography model includes:

Tuning the learning rate of the constructed image steganography network model, the number of images to be processed in batches, the iteration cycle and the optimizer.
The method for color image steganography based on generative adversarial networks according to claim 1, wherein the performance test of the image steganography model after parameter tuning includes:

Input the carrier image and the secret image in the test set into the image steganography network model after parameter tuning respectively, and output the steganographic image and the extracted image corresponding to the test set;

Carrying out a quality test on the steganographic image and the extracted image corresponding to the output test set;

Performing a security test on the steganographic image corresponding to the output test set;

The generalization ability test is carried out on the image steganography model after parameter tuning.
The color image steganography method based on generating an adversarial network according to claim 9, wherein the steganographic image and the extracted image corresponding to the output test set are subjected to quality testing, including:

Calculate the output PSNR value and SSIM value between the steganographic image corresponding to the test set and the cover image in the test set, and calculate and output the extracted image corresponding to the test set and the secret image in the test set PSNR value and SSIM value between;

According to the calculated PSNR value and SSIM value, the quality test result is obtained.