Strong-robustness image steganography method based on adversarial optimization and feedback connection

Through a strong robust image steganography method that optimizes the connection with feedback, using reversible neural network and feedback connection blocks, integrating simulated attack layer and dual enhancement module, combining differential consistency loss function, the robustness and The problem of insufficient security achieves high-quality image hiding and reconstruction, resists noise and JPEG compression, and reduces the risk of steganography analysis.

CN118921429APending Publication Date: 2024-11-08GUANGXI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410924659.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

When the existing robust image steganography method improves robustness, the hidden capacity and security are limited, making it difficult to effectively resist interference such as noise and JPEG compression during information transmission, and is easily detected by the steganography analysis model.

Method used

Adopt a robust image steganography method based on adversarial optimization and feedback connection. Through the reversible neural network design and feedback connection block, the simulated attack layer and dual enhancement module are integrated, and the differential consistency loss function is combined to improve the robustness of image steganography. and security.

Benefits of technology

The image quality and reconstruction effect are improved within the same training time, the resistance to noise and JPEG compression is enhanced, the detection risk of steganography analysis model is reduced, and the detection capacity is achieved is higher hidden capacity and better generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118921429A_ABST
    Figure CN118921429A_ABST
Patent Text Reader

Abstract

The invention discloses a strong-robustness image steganography method based on adversarial optimization and feedback connection, and mainly solves the problem that hiding capacity and security are weakened in order to improve robustness in the steganography process of most existing methods. The image steganography method provided by the invention comprises the following steps: 1) a carrier image and a secret image enter a hidden network, and a secret-containing image and loss information are obtained through a reversible block; 2) the secret-containing image passes through a simulation attack layer to obtain an attacked secret-containing image, and the network is subjected to confrontation optimization; (3) the attacked secret-containing image is restored to be in a non-attacked state through a content enhancement module and enters a reconstruction network along with the auxiliary variable, and a reconstructed secret image and a reconstructed carrier image are obtained; and 4) the reconstructed secret image is visually closer to the original secret image through the visual enhancement module.The steganography robustness is improved by the simulation attack layer and the double enhancement module, the security is enhanced by the proposed difference consistency loss, and meanwhile, the large hiding capacity is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information security, and in particular to a strong robust image steganography method based on adversarial optimization and feedback connection. Background Art

[0002] The rapid development of network technology has brought many conveniences to daily life. People can use various communication software to send information and receive various news and information on various social platforms. However, these network technologies also facilitate some attackers with ulterior motives. During the information transmission process, these attackers will monitor and steal the information. How to ensure the security of transmitted information during information communication has become an issue of increasing concern to people.

[0003] Steganography uses the insensitivity of human senses (such as vision and hearing) and the information redundancy of multimedia carriers to hide secret information. The information to be conveyed is embedded in multimedia carriers (such as text, images, audio, etc.), while the original carrier information can be kept relatively unchanged to achieve the effect of "steganography". Images are one of the multimedia carriers widely used in people's daily lives, and images contain rich information, including color, texture, shape, etc., which makes hiding secret information in images more flexible and difficult to detect. Therefore, image steganography has also become one of the research directions that has received much attention in information security.

[0004] The robustness of image steganography refers to the ability to effectively reconstruct the secret image after the secret image has been subjected to external attacks or image processing operations. In the actual information transmission process, the image may undergo various transformations and processing, such as compression, adding noise, etc. If the steganography method lacks robustness, the hidden secret information can be easily lost or destroyed during the image processing process, resulting in the unavailability or leakage of information. Most of the existing robust image steganography methods integrate simulated attacks in the steganography process and improve the robustness of the method through adversarial training. However, the improvement of robustness also limits the improvement of hiding capacity and security. Summary of the invention

[0005] The present invention provides a strong and robust image steganography method based on adversarial optimization and feedback connection. The method has achieved better performance than existing advanced steganography methods in terms of security and robustness, and can also hide color images of the same size, and the hiding capacity is also improved.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The robust image steganography method based on adversarial optimization and feedback connection includes the following steps:

[0008] 1) The carrier image and the secret image enter the hidden network designed based on the reversible neural network, and the secret image and loss information are obtained through N reversible blocks. The reversible block is composed of a radial coupling layer, and the feedback connection block is used as a learning module in the affine transformation;

[0009] 2) In the steganographic process, a simulated attack layer is integrated, and the secret image is passed through the simulated attack layer to obtain the attacked secret image, thereby performing adversarial optimization on network training;

[0010] 3) The attacked secret image is restored to its unattacked state through the content enhancement module, and the data sampled from the standard Gaussian distribution enters the reconstruction network as auxiliary variables to obtain the reconstructed secret image and the reconstructed carrier image;

[0011] 4) The reconstructed secret image is visually closer to the original secret image after the visual enhancement module;

[0012] 5) The optimization strategy for the overall network is divided into the design of training strategy and loss function.

[0013] The present invention further illustrates that a new learning module is proposed in step 1), and a feedback connection block is used to enable the deep convolution block to access the shallow features by adopting a skip connection method. After the feedback mechanism is introduced, the network uses the previous input information to dynamically adjust the current processing method, thereby reducing the sensitivity to interference and improving the overall robustness.

[0014] The present invention further illustrates that the simulated attack layer in step 2) takes into account common noise and JPEG compression. In the quantization operation of JPEG compression, the rounding function is lossy and non-differentiable, and the original JPEG algorithm cannot be used in network training. Therefore, an approximate rounding function is used instead to achieve simulation of JPEG compression.

[0015]

[0016] Where I is the image after DCT transformation divided by the quantization matrix, Q(I) is the result of approximate rounding of the image in the quantization operation of JPEG compression, and K is a balance parameter. The higher the value, the closer the effect is to the original rounding function, but the training time will also increase.

[0017] The present invention further illustrates that the dual enhancement modules in steps 3) and 4) include a content enhancement module and a visual enhancement module, wherein the content enhancement module restores the disturbed secret image to the original secret image as much as possible, thereby improving the quality of the reconstructed secret image; and the visual enhancement module makes the reconstructed secret image visually closer to the original secret image.

[0018] The present invention further illustrates that the design of the loss function in the step 5) proposes a difference consistency loss. The difference image is the result of directly subtracting two images. If the two images are more similar and the pixel values ​​are more identical, the pixel value of the obtained difference image should be 0, which is a completely black color in the image representation. Therefore, the present invention proposes a difference consistency loss, which is performed by performing L2 regularization on the difference image obtained by subtracting the secret image from the carrier image and the same-sized image with all pixel values ​​of 0, thereby reducing the pixel value of the difference image and constraining the difference image to be close to the completely black image. This can blur the visual information of the image, making it impossible for the attacker to intuitively observe the hidden location or content of the secret information, and can also reduce the probability of being detected by the steganalysis model, effectively reducing the risk of secret information leakage. It is defined as follows:

[0019]

[0020] Among them, X stego Indicates a secret image, X cover represents the carrier image, X black Represents an image with the same size as the difference image and all pixel values ​​are 0; X stego -X cover is a simple pixel difference; ||X stego -X cover |-X black | 2 is the distance between the difference images, using the mean squared error; the subscript 2 is the Euclidean norm.

[0021] Description of the technical innovation of the present invention:

[0022] A strong and robust image steganography method based on adversarial optimization and feedback connection. The present invention uses a reversible neural network to construct an image hiding and reconstruction network, and proposes a new feedback connection learning module. Compared with the commonly used learning module, it can make the image quality generated by the hiding and reconstruction network higher under the same training time. At the same time, in order to enhance the robustness of the image steganography method, a simulated attack layer is integrated in the steganography process, and the quality of the reconstructed secret image is iteratively optimized through the content enhancement module and the visual enhancement module constructed based on the feedback connection learning module, combined with the corresponding content loss function and visual loss function. In addition, in order to improve the network's ability to resist steganalysis, a difference consistency loss function is proposed, which slightly sacrifices the quality of the secret image in exchange for higher resistance to steganalysis. Experiments show that this method can hide a color secret image of the same size in a color carrier image, and can still reconstruct the secret image with high quality after interference attacks such as different noises and different intensities of JPEG compression, and can also effectively resist the detection of the steganalysis model based on deep learning.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] 1. In terms of image quality, better hiding and reconstruction effects are achieved within the same training time.

[0025] 2. In terms of robustness, the dual enhancement module constructed through the feedback connection module effectively reconstructs the secret image from the attacked confidential image with high quality.

[0026] 3. In terms of security, a difference consistency loss function is proposed to improve the resistance of the steganography method to the deep learning-based steganalysis model.

[0027] 4. The steganographic method proposed in the present invention has good generalization ability, can relatively hide larger capacity information, and still shows high robustness under attacks of various noises and interferences such as JPEG compression, and can effectively resist steganalysis detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the network structure of the enhancement module;

[0029] Figure 2 The present invention demonstrates the hiding and reconstruction effects of images;

[0030] Figure 3 The present invention is an implementation flow chart of an embodiment. DETAILED DESCRIPTION

[0031] The present invention is further described below in conjunction with the accompanying drawings and embodiments:

[0032] Figure 1 It is a schematic diagram of the network structure of the enhancement module;

[0033] Figure 2 The present invention demonstrates the hiding and reconstruction effects of images;

[0034] Figure 3 The present invention is an implementation flow chart of an embodiment.

[0035] Example:

[0036] The robust image steganography method based on adversarial optimization and feedback connection includes the following steps:

[0037] First, the carrier image X cover and secret image X secret The discrete wavelet transform (DWT) is performed simultaneously, and then the two enter the hidden network. After processing through N reversible blocks, the encrypted image X is obtained. stegoIn order to enhance the network's ability to adapt to different interferences, the encrypted image will pass through the simulated attack layer to obtain the attacked encrypted image The attack introduces unpredictable changes, disrupting the location and method of the secret image hiding, making it impossible for the reconstruction network to accurately reconstruct the secret image. Therefore, the attacked secret image will first be restored to its original state as much as possible through the content enhancement module. stego After discrete wavelet transformation together with the auxiliary variables sampled from the standard Gaussian distribution, the feedback is sent to the reconstruction network, and the reconstructed secret image X is obtained after the same processing of N reversible blocks. rse and reconstruct the carrier image X rco In order to further improve the quality of the reconstructed secret image, it is visually enhanced by a visual enhancement module to obtain an enhanced reconstructed secret image EX res .

[0038] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, after understanding the content and principle of the present invention, professionals in the field may make various modifications and changes in form and details without departing from the principle and structure of the present invention, but these modifications and changes based on the idea of ​​the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A strong and robust image steganography method based on adversarial optimization and feedback connection, characterized in that: The following steps are involved: 1) The carrier image and the secret image enter the hidden network designed based on the reversible neural network, and the secret image and loss information are obtained through N reversible blocks. The reversible block is composed of a radial coupling layer, and the feedback connection block is used as a learning module in the affine transformation; 2) In the steganographic process, a simulated attack layer is integrated, and the secret image is passed through the simulated attack layer to obtain the attacked secret image, thereby performing adversarial optimization on network training; 3) The attacked secret image is restored to its unattacked state through the content enhancement module, and the data sampled from the standard Gaussian distribution enters the reconstruction network as auxiliary variables to obtain the reconstructed secret image and the reconstructed carrier image; 4) The reconstructed secret image is visually closer to the original secret image after the visual enhancement module; 5) The optimization strategy for the overall network is divided into the design of training strategy and loss function.

2. The strong robust image steganography method based on adversarial optimization and feedback connection according to claim 1 is characterized in that: In the step 1), a new learning module is proposed, a feedback connection block is used, and a jump connection method is adopted to enable the deep convolution block to access the shallow features. After the feedback mechanism is introduced, the network uses the previous input information to dynamically adjust the current processing method, thereby reducing the sensitivity to interference and improving the overall robustness.

3. The strong robust image steganography method based on adversarial optimization and feedback connection according to claim 1, characterized in that: The simulated attack layer in step 2) considers common noise and JPEG compression. In the quantization operation of JPEG compression, the rounding function is lossy and non-differentiable. The original JPEG algorithm cannot be used in network training. An approximate rounding function is used as a substitute to achieve the simulation of JPEG compression: Where I is the image after DCT transformation divided by the quantization matrix, Q(I) is the result of approximate rounding of the image in the quantization operation of JPEG compression, and K is a balance parameter. The higher the value, the closer the effect is to the original rounding function, but the training time will also increase.

4. The strong robust image steganography method based on adversarial optimization and feedback connection according to claim 1, characterized in that: The dual enhancement module in steps 3) and 4) includes a content enhancement module and a visual enhancement module, wherein the content enhancement module restores the disturbed secret image to the original secret image as much as possible, thereby improving the quality of the reconstructed secret image; The visual enhancement module makes the reconstructed secret image visually closer to the original secret image.

5. The strong robust image steganography method based on adversarial optimization and feedback connection according to claim 1, characterized in that: The optimization strategy in the step 5) is that the whole network adopts three-step training; in the first step, only the hidden network and the reconstruction network are trained, and the obtained secret image does not pass through the simulated attack layer; in the second step, the obtained secret image passes through the simulated attack layer, and starts to train the content enhancement module and the visual enhancement module, but freezes the parameters of the hidden network and the reconstruction network. In addition, in order to make the second step more stable, the total number of training cycles in the second step is divided into several sections, and the secret image will be fixedly attacked by a certain interference in a certain section, and will be fixedly attacked by another interference in the next section; in the third step, the hidden network and the reconstruction network are retrained, and the generated secret image will be randomly attacked by a certain interference. The secret image after the attack is enhanced by the content enhancement module, and then input into the reconstruction network for secret image reconstruction, and finally the reconstructed secret image will enter the visual enhancement module for visual enhancement.

6. The strong robust image steganography method based on adversarial optimization and feedback connection according to claim 1, characterized in that: The design of the loss function in step 5) proposes a difference consistency loss; the difference image is the result of directly subtracting two images. If the two images are more similar and the pixel values ​​are more identical, the pixel value of the obtained difference image should be 0, which is a completely black color in the image representation; the difference consistency loss performs L2 regularization on the difference image obtained by subtracting the secret image from the carrier image and the same-sized image with all pixel values ​​of 0, reduces the pixel value of the difference image, constrains the difference image to be close to the completely black image, blurs the visual information of the image, and makes it impossible for the attacker to intuitively observe the hidden position or content of the secret information. It can also reduce the probability of being detected by the steganalysis model, effectively reducing the risk of secret information leakage; the difference consistency loss function is defined as follows: Among them, X stego Indicates a secret image, X cover represents the carrier image, X black Represents an image with the same size as the difference image and all pixel values ​​are 0; X stego -X cover is a simple pixel difference; ||X stego -X cover |-X black | 2 is the distance between the difference images, using the mean squared error; the subscript 2 is the Euclidean norm.

Citation Information

Cited By

  • Anti-interference perception encryption method for unmanned aerial vehicle image

    CN122293803A