Method for realizing completely reversible image steganography based on reversible neural network

Abstract

The invention discloses a completely reversible image steganography method based on a reversible neural network, and aims to improve the robustness, capacity and safety of image steganography. The method comprises the following steps: 1) obtaining a processed secret image S through N reversible blocks, and adding a carrier image and the processed secret image S; 2) in a steganography process, integrating a simulation attack layer, and obtaining an attacked secret-containing image by the secret-containing image through the simulation attack layer, so as to carry out confrontation optimization on network training; (3) the attacked secret-containing image is restored to be in a non-attacked state through a content enhancement module, a cover image is found in a receiver database, and a processed secret image S is obtained through subtraction of the cover image and the secret-carrying image; 4) inputting the processed secret image S into a reversible neural network to recover a secret image; and 5) the optimization strategy of the whole network is divided into the design of a training strategy and a loss function. Through the innovative architecture of the invention, the complete reversibility of image hiding and reconstruction is successfully realized.

Description

Technical Field

[0001] The invention belongs to the field of information security, and in particular relates to a method for realizing fully reversible image steganography based on a reversible neural network. Background Art

[0002] The rapid development of network technology has brought great convenience to people's daily lives. People can transmit information through various communication software or get news information on social platforms. However, the popularity of these technologies has also provided convenience for some malicious attackers, who can monitor and steal information during the transmission process, and information security issues have therefore attracted more and more attention. How to effectively ensure the security of transmission content during information communication has become one of the core issues in today's information society.

[0003] Steganography is a technology that uses the insensitivity of human senses and the information redundancy of multimedia carriers to hide secret information. By embedding secret information into multimedia carriers (such as text, images, audio, etc.) while keeping the appearance or content of the original carrier basically unchanged, the purpose of "secret transmission" is achieved. Among various multimedia carriers, images are an ideal choice for steganography because of their rich information and multidimensional features such as color, texture and shape. Image steganography has become a research direction that has attracted much attention in the field of information security due to its flexibility and concealment.

[0004] The robustness of image steganography refers to the ability to effectively reconstruct secret information after the secret image has been subjected to external attacks or image processing operations (such as compression, noise interference, etc.). In actual communication scenarios, images may undergo a variety of transformations and processing operations, such as JPEG compression, adding noise, etc. If the steganography method lacks sufficient robustness, the hidden secret information can easily be destroyed or lost in these processes, resulting in the information being unable to be recovered or leaked. At present, existing robust image steganography methods usually enhance robustness by integrating simulated attacks and combining adversarial training. However, the improvement of robustness is usually accompanied by a reduction in hidden capacity and security. Therefore, how to improve the capacity and security of information hiding while ensuring robustness remains a technical problem that needs to be solved urgently. Summary of the invention

[0005] The present invention provides a method for realizing fully reversible image steganography based on a reversible neural network. The method has achieved performance superior to existing advanced steganography methods in terms of security, robustness and image fidelity, and can also hide color images of the same size, and the hiding capacity is also improved. In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] 1. A method for realizing fully reversible image steganography based on a reversible neural network, characterized in that it comprises the following steps:

[0007] 1) First, the secret image enters the hidden network designed based on the reversible neural network, and the processed secret image S is obtained through N reversible blocks. The carrier image and the processed secret image S are added. The reversible block is composed of a radiation coupling layer;

[0008] 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;

[0009] 3) The attacked secret image is restored to its unattacked state by the content enhancement module, and the corresponding cover image is found in the recipient's database. The processed secret image S is obtained by subtracting the cover image from the secret image;

[0010] 4) Input the processed secret image S into the N inverse processes of the reversible neural network to recover the complete secret image;

[0011] 5) The optimization strategy for the overall network is divided into the design of training strategy and loss function.

[0012] The present invention further illustrates that an innovative architecture design is introduced in the step 1), the core concept of which is to input the secret image into the reversible neural network only during the forward transfer process, and after the processing is completed, the secret image processed by the neural network is superimposed on the secret image. This design significantly simplifies the process of information embedding and improves the confidentiality of information hiding. On this basis, the present invention adds a deep convolution block to the reversible module, aiming to enhance the network's sensitivity to small changes in the image through deep feature extraction, thereby playing a key role in reducing the impact of external interference. The deep convolution block can adaptively adjust the size of the convolution kernel and the feature extraction method, and then dynamically optimize the current processing process, so that the system can more efficiently cope with various complex image processing operations or external attacks. At the same time, the deep convolution block also improves the ability to capture subtle features in the secret image through multi-scale feature fusion, so that the image information can be hidden under the premise of high fidelity. This design effectively improves the robustness of the image steganography method and the accuracy of image reconstruction, and provides a reliable guarantee for secure information transmission.

[0013] The present invention further illustrates that in the step 2), the simulated attack layer is an important component of the system for enhancing robustness, and common attack types are mainly considered, including noise, JPEG compression and random cropping. In order to ensure the wide applicability of the system in practical applications, the present invention has adjusted the parameters of noise intensity, JPEG compression intensity and random cropping size through multiple experiments. These experiments help the present invention determine the optimal performance of the network under different intensities. Specifically, for the processing of noise, the present invention tests multiple types such as Gaussian noise and salt and pepper noise, gradually increases its intensity, and trains the network to adapt to these changes; for JPEG compression, the present invention focuses on the impact of different compression ratios on image quality and steganographic effect, and by adjusting the compression ratio, the system can effectively retain secret information under low to high compression ratio conditions; for random cropping, the present invention controls the randomness of cropping size and cropping position to ensure that the network can still successfully reconstruct the secret image under various image cropping conditions. This process improves the robustness of the system in the face of various image processing or attack operations, and ensures the robustness of information hiding.

[0014] The present invention further illustrates that in the step 3), the content enhancement module is one of the key components for improving the image steganography performance. The main function of this module is to restore the secret image to a state close to the original secret image as much as possible in view of the influence of external interference (such as noise, compression, cropping, etc.) on the secret image. Through the processing of the content enhancement module, the system can effectively improve the quality of the reconstructed secret image and ensure the accuracy of information transmission. Especially in the case of random cropping, part of the information of the image may be lost, resulting in the inability to fully restore the secret information. In order to solve this problem, the present invention proposes a processing method based on neighborhood pixel filling. The method analyzes the neighborhood pixel features of the cropped area and uses interpolation or convolution operations to fill the missing parts, thereby restoring the original image as much as possible. This method can not only significantly reduce the impact of cropping on image quality, but also improve the availability of the secret image and ensure high-precision reconstruction of information.

[0015] The present invention further illustrates that the optimization strategy in step 5) is the key to improving the performance of the entire system, and includes three training stages, which are respectively optimized and trained for the hidden network, the reconstruction network, and the content enhancement module. In the first stage, the hidden network and the reconstruction network are trained with emphasis. The secret images in this stage are not processed by the simulated attack layer, and the purpose is to ensure that the system can achieve efficient information hiding and reconstruction under interference-free conditions. The training in this stage lays the foundation for subsequent robustness enhancement. In the second stage, the secret images are processed by the simulated attack layer, and the focus of training is turned to the content enhancement module, while the parameters of the hidden network and the reconstruction network remain frozen. In this stage, in order to improve the stability of training and the generalization ability of the model, the present invention divides the total training cycle into several small cycles, and in each small cycle, the secret image is fixedly subjected to a certain type of interference attack, such as noise or compression. In the next small cycle, the attack type will change. Through this staged training method, the system can gradually adapt to different types of external interference and enhance its robustness. In the third stage, the hidden network and the reconstruction network are unfrozen again, and the secret image is randomly subjected to a certain interference attack. The attacked secret image will be input into the content enhancement module for processing, and then passed to the reconstruction network to complete the reconstruction of the secret image. This three-stage training method not only effectively improves the robustness of image steganography, but also ensures high-quality information transmission and reconstruction.

[0016] Description of the technical innovation of the present invention:

[0017] The present invention proposes an image hiding and reconstruction method based on a reversible neural network. By constructing a hiding and reconstruction network, efficient steganalysis and recovery of image information are achieved. Compared with the traditional steganalysis architecture, the present method designs an innovative model hiding architecture, which can generate higher quality hidden images and reconstructed images within the same training time. In order to further improve the robustness of the steganalysis process, the system integrates a simulated attack layer, and enhances the anti-attack ability of the system by adding common interferences such as noise and JPEG compression. In addition, the content enhancement module is combined with a specific loss function, and iterative optimization is performed during the steganalysis process to ensure that the quality of the reconstructed secret image reaches the best. In order to improve the network's ability to resist steganalysis, the present invention also proposes a method for cleverly superimposing a secret image with a carrier image, so that the system exhibits stronger resistance to steganalysis. Experimental results show that the method can not only efficiently hide color secret images of the same size in color carrier images, but also successfully reconstruct secret images with high quality after interferences such as different noises and different intensities of JPEG compression. At the same time, the system can effectively resist the detection of a steganalysis model based on deep learning, and has significant advantages in both information security and robustness.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] 1. Compared with the existing methods, the present invention has achieved significant improvements in image hiding and reconstruction effects, and the generated images are of higher quality and better detail fidelity.

[0020] 2. By introducing a content enhancement module, the present invention can efficiently reconstruct a high-quality secret image from an attacked confidential image, and exhibits excellent robustness in the face of common interferences such as noise and compression.

[0021] 3. The present invention proposes an innovative secret image and encrypted image superposition strategy, which effectively improves the ability of the steganographic method to combat the deep learning-based steganalysis model and significantly enhances the security of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention demonstrates the hiding and reconstruction effects of images;

[0023] Figure 2 The present invention is an implementation flow chart of an embodiment. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with the accompanying drawings and embodiments:

[0025] Figure 1 The present invention demonstrates the hiding and reconstruction effects of images;

[0026] Figure 2 The present invention is an implementation flow chart of an embodiment.

[0027] Example:

[0028] The robust image steganography method based on adversarial optimization and feedback connection includes the following steps:

[0029] First, the secret image Xsecret is decomposed into different frequency components by using discrete wavelet transform (DWT). In this way, the present invention can divide the image into high-frequency and low-frequency domains, so that the secret image can be effectively and quickly hidden in the high-frequency domain. Embedding secret information in the high-frequency domain helps to better resist the influence of noise. It should be noted that only the decomposed secret image xsecret is input into the reversible neural network to generate the encoded secret image se. In order to restore the secret image Se from the high-frequency and low-frequency domains after the reversible neural network, the inverse transform of DWT, namely the inverse discrete wavelet transform (IWT), is used. There are two methods to hide Se into a random cover image, namely 1) Xcover-Se=Xstego, 2) Xcover+Se=Xstego, to generate a secret image. Subsequently, the secret image is sent to the receiver, and is interfered by Gaussian noise, JPEG and dropout on the way. After receiving the secret image, the receiver inputs it into the image enhancement module for enhancement. Then, the corresponding cover image is found through the database of cover images owned by both the sender and the receiver. Extract the encoded secret image Se from the secret image, corresponding to the above two methods, namely 1) Xcover-Xstego=Se, 2) Xstego-Xcover=Se. After Se is inversely transformed by DWT, it is input into the inverse process of the reversible neural network, and finally the secret image Xsecret is restored.

[0030] 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 fully reversible image steganography method based on a reversible neural network, characterized in that: The following steps are involved: 1) First, the secret image enters the hidden network designed based on the reversible neural network, and the processed secret image S is obtained through N reversible blocks. The carrier image and the processed secret image S are added. The reversible block is composed of a radiation coupling layer; 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 by the content enhancement module, and the corresponding cover image is found in the recipient's database. The processed secret image S is obtained by subtracting the cover image from the secret image; 4) Input the processed secret image S into the N inverse processes of the reversible neural network to recover the complete secret image; 5) The optimization strategy for the overall network is divided into the design of training strategy and loss function.

2. According to claim 1, a method for realizing fully reversible image steganography based on a reversible neural network is characterized in that: A new architecture design is introduced in step 1); this design only inputs the secret image into the forward process of the reversible neural network and superimposes the processed secret image into the secret image; to this end, we introduce deep convolution blocks in the reversible module and dynamically adjust the current processing method, thereby effectively reducing the sensitivity to external interference and improving the robustness and high fidelity of the image.

3. According to claim 1, a method for realizing fully reversible image steganography based on a reversible neural network is characterized in that: The simulated attack layer involved in step 2) takes into account common noise, JPEG compression, and random cropping. We continuously adjust the parameters of noise intensity, JPEG compression intensity, and cropping size through experiments to ensure that the network can achieve the best performance when facing noise of different intensities and random cropping of different sizes.

4. According to claim 1, a method for realizing fully reversible image steganography based on a reversible neural network is characterized in that: The content enhancement module in step 3) can restore the disturbed secret image to the original secret image as much as possible, thereby improving the quality of the reconstructed secret image; for the random cropping problem, we introduced a processing method based on neighborhood pixel filling.

5. According to claim 1, a method for realizing fully reversible image steganography based on a reversible neural network is characterized in that: The optimization strategy in step 5) is divided into three training stages; in the first stage, only the hidden network and the reconstruction network are trained, and the generated secret image does not pass through the simulated attack layer; in the second stage, the generated secret image passes through the simulated attack layer, trains the content enhancement module, and freezes the parameters of the hidden network and the reconstruction network; in order to improve the stability of the second stage, the total training cycle is divided into several stages, and each stage applies a specific interference attack; in the third stage, the hidden network and the reconstruction network are retrained, and the generated secret image will randomly suffer from different types of interference attacks. After being processed by the content enhancement module, it is input into the reconstruction network to complete the reconstruction of the secret image.

Images