A dual-branch generation algorithm to counter watermark attacks

Abstract

The present invention discloses a dual-branch generative adversarial watermark attack algorithm, which relates to the field of image processing technology and is characterized in that it comprises the following steps: S1: constructing a precise two-branch generative adversarial network architecture, including a detail-preserving generative adversarial network and a watermark attack generative adversarial network; S2: image difference analysis: using a precise two-branch generative adversarial network architecture to perform pixel-by-pixel difference operations between the original image and the watermarked image to extract image details and watermark information, and outputting two independent channels, namely, an image detail map and a watermark information map; S3: generating a detail-preserving image. The present invention can effectively counter existing watermarking technologies while maintaining image quality, provides a new tool for watermark security assessment, and promotes the advancement of watermarking technology.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a dual-branch generation algorithm for resisting watermark attacks. Background Art

[0002] Digital watermarking technology is an advanced information hiding method that aims to embed specific identification information into digital media content to achieve purposes such as copyright protection, content authentication, and hidden communication. This technology embeds imperceptible watermarks in image, audio, or video data, so that the media content can be traced back to the original owner or its integrity can be verified when it is subjected to unauthorized use, tampering, or distribution. Digital watermarks can be divided into visible watermarks and invisible watermarks, among which invisible watermarks are embedded through spatial or frequency domain methods to ensure the concealment of the watermark and the robustness to signal processing operations. Although digital watermarking technology plays an important role in improving media security, it also faces challenges such as how to balance the invisibility, robustness, and security of the watermark. With the development of technology, researchers continue to explore new watermark embedding algorithms to enhance its resistance to attacks while maintaining minimal impact on the original media content.

[0003] Generative adversarial network (GAN) is a cutting-edge deep learning model, which consists of two key components: generator and discriminator. It has achieved significant improvement in image generation and discrimination capabilities through adversarial training. The generator is responsible for creating realistic images, while the discriminator is dedicated to distinguishing real images from generated images. The two compete with each other during the training process, driving the continuous improvement of the quality of generated images. The network structure of GAN is usually based on convolutional neural networks, which makes it show great potential in image processing fields such as artistic creation, data enhancement and image editing. However, the training process of GAN may face challenges of stability and mode collapse, requiring careful design of network architecture and training strategies. In addition, evaluating the quality of generated images is also more complicated than traditional image processing, requiring comprehensive consideration of visual authenticity, diversity and feedback from adversarial discriminators. Despite the challenges, GAN has broad application prospects in image generation and editing, and continues to promote the development of the field of artificial intelligence.

[0004] Image processing and analysis play a vital role in watermark attacks. First, image processing technology is used to optimize the embedding process of watermarks, ensure the concealment of watermarks, and make them difficult to be visually detected. At the same time, the robustness of watermarks is improved by image enhancement and filtering, ensuring that the watermark information can still be accurately extracted after the image is processed by common processing operations. Secondly, image analysis technology not only helps to evaluate the stability and resistance of watermarks, but also is used to detect and identify whether an image has been attacked by watermarks, as well as the type and extent of the attack. In addition, the development of deep learning, especially the application of generative adversarial networks (GANs), has provided new tools for watermark attacks. Attackers can use GANs to generate watermarks that are difficult to detect, while defenders can use these technologies to improve the resistance of watermarks. Image quality assessment is also key, which ensures the practicality of watermark technology, that is, while protecting copyright and integrity, the visual quality of the image is not sacrificed. The comprehensive application of these technologies has promoted the continuous evolution of watermark attack and defense strategies.

[0005] Watermark attack technology plays a vital role in promoting the development of watermark technology. First, it promotes the continuous progress and innovation of watermark technology by revealing the potential weaknesses of existing watermark algorithms and inspiring researchers to develop safer and more robust solutions. Second, as an efficient security testing method, watermark attacks help developers and researchers to deeply understand the attack scenarios that may be encountered in practical applications, which is crucial for designing watermark systems that can resist complex attacks. In addition, the development of watermark attack technology provides an effective testing tool for evaluating the robustness of watermark algorithms. By simulating different types of attacks, the stability and reliability of watermark algorithms in the face of real-world threats can be accurately tested. This not only strengthens the security of watermark technology, but also provides a scientific basis for the standardized evaluation and testing of watermark algorithms.

[0006] Despite the continuous progress of watermark attack technologies, they still face a series of challenges. Traditional methods often damage the quality of the media and cause distortion when performing operations such as image transformation, resizing, and compression on digital media, which not only affects the user experience, but also reduces the value of the watermarked image in practical applications. Although the new generation of watermark attack technologies can circumvent existing watermark detection mechanisms, they are still in their infancy and face defects such as insufficient attack concealment, limited effectiveness, and overly simplified network structure. The existing technologies have the following shortcomings:

[0007] 1. Insufficient watermark attack capabilities: Existing watermark attack technologies are often unable to effectively destroy or remove the watermark information embedded in the image, resulting in the watermark still being detectable in the image after the attack, and the attacker’s purpose cannot be achieved.

[0008] 2. Poor imperceptibility of the attacked image: Traditional attack methods may damage the image quality in the process of removing the watermark, resulting in a large visual difference between the attacked image and the original image, which cannot meet the requirements of conducting an effective attack without reducing the image quality.

[0009] 3. Removing watermarks while retaining image details: During the watermark attack process, how to effectively remove the watermark information while maintaining the details and quality of the image, ensuring that the image after the attack is still practical and has acceptable visual quality. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a dual-branch generation anti-watermark attack algorithm, which can effectively fight against existing watermark technology while maintaining image quality, provide a new tool for watermark security assessment, and promote the advancement of watermark technology.

[0011] The present invention adopts the following technical solutions to achieve the invention objectives:

[0012] A dual-branch generation algorithm for resisting watermark attacks, characterized by comprising the following steps:

[0013] S1: Construct a sophisticated two-branch GAN architecture, including a detail-preserving GAN and a watermark-attack GAN;

[0014] S2: Image difference analysis: A sophisticated two-branch generative adversarial network architecture is used to perform pixel-by-pixel difference operations between the original image and the watermarked image to extract image details and watermark information. The output is two independent channels, namely the image detail map and the watermark information map.

[0015] Watermarked image WI and original watermark information OI Subtract to get image details ;

[0016] (1)

[0017] Watermarked image WI and the original image O Subtract and get watermark information I ;

[0018] (2)

[0019] S3: Generate detail-preserving image: The image detail map is then input into the detail-preserving generative adversarial network, which generates a detail-preserving image;

[0020] S4: Generate watermark-destroyed image: watermark information map and corresponding inverted watermark information Picture oneThe same input is fed into the watermark attack generative adversarial network to generate an image with the watermark destroyed;

[0021] S5: Image fusion: The detail-preserving image and the watermark-destroyed image are added pixel-wise to produce the final post-attack image.

[0022] As a further limitation of the technical solution, S3 includes: in the process of generating adversarial, designing a generator G , which maps the image details As input, it generates an image , the image not only retains the features of the detail image, but also should be close to the original image ; At the same time, design a discriminator , distinguish the generated images and the original image ;

[0023] The following is the objective function of the generator:

[0024] (3)

[0025] in: It is a confrontational loss;

[0026] It is the loss of content;

[0027] It is the loss of detail;

[0028] is the target edge enhancement loss;

[0029] is a weight parameter used to balance the contribution of different loss terms;

[0030] The following is the objective function of the discriminator:

[0031] Discriminator The goal is to maximize the following function to correctly distinguish the original image from the generated image:

[0032] (4)

[0033] Among them: The first term is the loss of the original image, which encourages the discriminator For the original image O Give high ratings;

[0034] The second term is the loss of generated images, which encourages the discriminator For the generator G Generated Image Give a low rating.

[0035] As a further limitation of the present technical solution, the S4 includes: the watermark information is I , set the anti-watermark information to RI , the output is the image after watermark destruction AI , the formula for generating adversarial is as follows:

[0036] (5)

[0037] Generator G Receive two inputs, the watermark information is I And the anti-watermark information is RI ;

[0038] Set the anti-watermark information to RI is the image after counter-watermarking, and the generator G Through its parameters The watermark information is I The anti-watermark information is RI Learning and generating watermark-corrupted images AI .

[0039] As a further limitation of the present technical solution, the S5 includes: AI and detail preservation map Fusion into post-attack image WIAA , which can be achieved by fusion function F To achieve this, this function will destroy the image after the watermark AI and detail-preserving images DPI As input, and output the fused image WIAA , this process can be expressed by the following formula:

[0040] (6)

[0041] in: F is the fusion function;

[0042] are the deep learning model parameters.

[0043] As a further limitation of the present technical solution, the detail-preserving generative adversarial network consists of a detail-preserving generator and a detail-preserving discriminator, which work together to generate high-quality detail-preserving images.

[0044] As a further limitation of the technical solution, the specific process of S3 is as follows:

[0045] S31: Detail extraction by detail-preserving generator: The detail-preserving generator receives the image detail map as input. The goal of the detail-preserving generator is to use the detail information to generate a detail-preserving image that is visually similar to the original image while removing the watermark information.

[0046] S32: authenticity judgment of detail-preserving discriminator: the generated detail-preserving image is input to the detail-preserving discriminator together with the original image, and the detail-preserving discriminator provides feedback to guide the detail-preserving generator to optimize the generated detail-preserving image so that it cannot be distinguished from the original image visually;

[0047] S33: Adversarial training: The detail-preserving generator and detail-preserving discriminator compete with each other in the adversarial training process;

[0048] S34: Iterative optimization of detail-preserving generator: Based on the feedback from the detail-preserving discriminator, the detail-preserving generator adjusts internal parameters through back-propagation and gradient descent optimization algorithms to generate more realistic detail-preserving images, until the images generated by the detail-preserving generator can effectively remove watermark information without sacrificing visual quality;

[0049] S35: Result verification: Finally, the generated detail-preserving image is evaluated by the detail-preserving discriminator. If it can frequently deceive the detail-preserving discriminator, that is, it is judged as a real image, then it is considered that the detail-preserving generator has successfully completed the task.

[0050] As a further limitation of the present technical solution, the watermark attack generation adversarial network includes a watermark attack generator and a watermark attack discriminator.

[0051] As a further limitation of the present technical solution, the specific process of S4 is as follows:

[0052] S41: Initialization of watermark attack generator: The watermark attack generator receives the original watermark information as input;

[0053] S42: Generation of damaged watermark information: The watermark attack generator processes the input original watermark information through its internal network structure to generate damaged watermark information;

[0054] S43: True or false judgment of watermark attack discriminator: The generated damaged watermark information is then input into the watermark attack discriminator together with the corresponding reversed watermark information. The watermark attack discriminator is responsible for evaluating the two types of information and making a true or false judgment;

[0055] S44: Adversarial training mechanism: The watermark attack generator and the watermark attack discriminator compete with each other in adversarial training;

[0056] S45: Feedback optimization of watermark attack generator: the true or false judgment result of the watermark attack discriminator is fed back to the watermark attack generator to guide the parameter update of the watermark attack generator;

[0057] S46: Output of damaged watermark information: After adversarial training and iterative optimization, the watermark attack generator can eventually generate watermark information that is visually similar to the reversed watermark information, so that it can deceive the watermark attack discriminator.

[0058] As a further limitation of the present technical solution, the reverse watermark information graph is a reverse representation of the watermark.

[0059] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention includes four core innovations: 1. Dual-branch structure design: The dual-branch structure design adopted by the network is one of its significant innovations. This design allows the network to decompose the watermark attack task into two subtasks: the preservation of image details and the destruction of watermark information. Each branch has its own dedicated generator and discriminator, so that the network can edit and watermark attack images more accurately, thereby improving the overall performance and effect. 2. Separate processing of details and watermark information: The network separates the image detail information from the watermark information so that the two types of information can be processed independently in different network branches. This separation processing strategy not only improves the targeting and effectiveness of the attack, but also allows the network to operate on the watermark information more accurately without compromising the image quality. 3. Adversarial training of generative adversarial networks: Using the adversarial training mechanism of generative adversarial networks (GANs), the generator and the discriminator compete with each other during the training process. The generator is committed to generating watermark-destroying images that can deceive the discriminator, while the discriminator continuously improves its recognition ability. This adversarial training not only enhances the generator's ability to generate high-quality images, but also makes the generated images more visually realistic and difficult to be recognized by existing watermark detection technologies. 4. Innovative watermark attack paradigm: The watermark attack paradigm proposed by the network is an innovative method that can not only verify the security of existing watermark technology, but also promote the development of watermark technology. By simulating the strategies that attackers may use, the network can reveal the potential weaknesses of watermark technology, thereby providing guidance for the improvement of watermark technology. The application of this method will help to build a more robust and reliable digital media protection mechanism. These four innovations together constitute the unique advantages of the dual-branch generation adversarial watermark attack network, giving it significant potential and application value in the field of watermark attacks. Through this design, the network can effectively counter existing watermark technology while maintaining image quality, providing a new tool for watermark security assessment and promoting the advancement of watermark technology. Through adversarial training, the generator of the present invention learns to generate images that are almost visually indistinguishable, which makes the attack results more difficult to be discovered by the detection system and improves the concealment of the attack. The detail preservation branch focuses on maintaining the visual quality of the image, ensuring that the details and texture information of the image are preserved to the greatest extent during the attack, reducing the distortion and noise that may be introduced by traditional attack methods. The watermark information attack branch specifically destroys the watermark information instead of rough processing the entire image, which makes the attack more precise and effectively reduces the possibility of accidentally damaging the non-watermark part. Using deep learning technology, especially generative adversarial networks, the network can learn complex image patterns and watermark embedding technology, providing a more advanced attack method. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Picture 1Generate an adversarial network flow chart for the two branches of the present invention.

[0061] Picture 2 The block diagram of the Generative Adversarial Network architecture is reserved for the details of the present invention.

[0062] Picture 3 The watermark attack generation adversarial network structure block diagram of the present invention.

[0063] Picture 4 This is a flow chart of obtaining reverse watermark information of the present invention. DETAILED DESCRIPTION

[0064] A specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation.

[0065] The present invention comprises the following steps:

[0066] S1: Construct a sophisticated two-branch Generative Adversarial Network (GAN) architecture specifically designed to perform watermark attack tasks while maintaining the visual quality of the original image, including Detail-Preserving GAN and Watermark-Attack GAN.

[0067] S2: Image difference analysis: Use a sophisticated two-branch generative adversarial network architecture to perform pixel-by-pixel difference operations between the original image and the watermarked image to extract image details and watermark information. The output is two independent channels, namely the image detail map and the watermark information map.

[0068] Watermarked image WI and original watermark information OI Subtract to get image details ;

[0069] (1)

[0070] Watermarked image WI and the original image O Subtract and get watermark information I ;

[0071] (2)

[0072] S3: Generate detail-preserving image: The image detail map is then fed into a detail-preserving generative adversarial network, which is designed to learn and preserve the visual details of the image while ignoring the watermark information as much as possible. The detail-preserving generative adversarial network generates a detail-preserving image that is visually similar to the original image, but the watermark information has been faded out.

[0073] The detail-preserving generative adversarial network consists of a detail-preserving generator and a detail-preserving discriminator, which work together to generate high-quality detail-preserving images.

[0074] In the process of generating adversarial, design a generator G , which maps the image details As input, it generates an image , the image not only retains the features of the detail image, but also should be close to the original image ; At the same time, design a discriminator , distinguish the generated images and the original image ;

[0075] The following is the objective function of the generator:

[0076] (3)

[0077] in: It is a confrontational loss;

[0078] It is the loss of content;

[0079] It is the loss of detail;

[0080] is the target edge enhancement loss;

[0081] is a weight parameter used to balance the contribution of different loss terms;

[0082] The following is the objective function of the discriminator:

[0083] Discriminator The goal is to maximize the following function to correctly distinguish the original image from the generated image:

[0084] (4)

[0085] Among them: The first term is the loss of the original image, which encourages the discriminator For the original image O Give high ratings;

[0086] The second term is the loss of generated images, which encourages the discriminator For the generator G Generated Image Give a low rating;

[0087] The specific process of S3 is as follows:

[0088] S31: Detail extraction by detail-preserving generator: The detail-preserving generator receives as input an image detail map, which is obtained by subtracting the original image from the watermarked image. The goal of the detail-preserving generator is to use this detail information to generate a detail-preserving image that is visually similar to the original image while removing the watermark information.

[0089] S32: Authenticity judgment of detail-preserving discriminator: The generated detail-preserving image is input to the detail-preserving discriminator together with the original image. The task of the detail-preserving discriminator is to evaluate the two images to distinguish whether they are original images. The detail-preserving discriminator provides feedback to guide the detail-preserving generator to optimize the generated detail-preserving image so that it cannot be visually distinguished from the original image.

[0090] S33: Adversarial training: The detail-preserving generator and detail-preserving discriminator compete with each other during adversarial training. The detail-preserving generator continuously learns how to generate detail-preserving images that can deceive the detail-preserving discriminator, while the detail-preserving discriminator continuously improves its recognition ability to distinguish real images from generated images.

[0091] S34: Iterative optimization of detail-preserving generator: Based on the feedback from the detail-preserving discriminator, the detail-preserving generator adjusts its internal parameters through optimization algorithms such as back-propagation and gradient descent to generate more realistic detail-preserving images. This process is iterative until the image produced by the detail-preserving generator can effectively remove the watermark information without sacrificing visual quality.

[0092] S35: Result verification: Finally, the generated detail-preserving image is evaluated by the detail-preserving discriminator. If it can frequently deceive the detail-preserving discriminator, that is, it is judged as a real image, then it is considered that the detail-preserving generator has successfully completed the task. The image obtained in this way is visually highly consistent with the original image, retains the image details, and meets the target requirements of the algorithm.

[0093] The entire algorithm flow is a dynamic learning and adaptation process, which requires fine-tuning of the network structure and training strategy to ensure that the generated detail-preserving images are visually indistinguishable from the original images, while achieving effective removal of watermark information.

[0094] The watermark attack generative adversarial network includes a watermark attack generator and a watermark attack discriminator, which focuses on generating damaged watermark information that can mislead the watermark detection system.

[0095] S4: Generate a watermark-destroyed image: The watermark information graph and the corresponding inverted watermark information graph (i.e., the reverse representation of the watermark) are input into the watermark attack generative adversarial network. The training goal of the network is to identify and effectively destroy the watermark information and generate a watermark-destroyed image.

[0096] The reverse watermark information graph is the reverse representation of the watermark, Picture 4 Process obtained.

[0097] The watermark information is I , set the anti-watermark information to RI , the output is the image after watermark destruction AI , the formula for generating adversarial is as follows:

[0098] (5)

[0099] Generator G Receive two inputs, the watermark information is I And the anti-watermark information is RI ;

[0100] Set the anti-watermark information to RI is the image after counter-watermarking, and the generator G Through its parameters The watermark information is I The anti-watermark information is RI Learning and generating watermark-corrupted images AI ;

[0101] Generator G It is a multi-layer neural network that learns how to learn to reverse the watermark information. RI features and incorporate them into the output;

[0102] The specific process of S4 is as follows:

[0103] S41: Initialization of watermark attack generator: The watermark attack generator receives original watermark information as input, the original watermark information is extracted from the watermarked image, and the purpose is to generate damaged watermark information;

[0104] S42: Generation of damaged watermark information: The watermark attack generator processes the input original watermark information through its internal network structure to generate damaged watermark information. This process simulates various attacks that the watermark may suffer, such as compression, filtering or geometric transformation, aiming to reduce the detectability of the watermark;

[0105] S43: True or false judgment of watermark attack discriminator: The generated damaged watermark information is then input into the watermark attack discriminator together with the corresponding inverted watermark information (i.e., logically opposite watermark information, used to simulate the idealized watermark destruction effect). The watermark attack discriminator is responsible for evaluating the two types of information and making a true or false judgment;

[0106] S44: Adversarial training mechanism: The watermark attack generator and the watermark attack discriminator compete with each other in adversarial training. The goal of the watermark attack generator is to generate damaged watermark information that can deceive the watermark attack discriminator, while the watermark attack discriminator strives to improve its recognition ability to distinguish between the reversed watermark and the fake watermark generated by the watermark attack generator;

[0107] S45: Feedback optimization of watermark attack generator: The true or false judgment result of the watermark attack discriminator is fed back to the watermark attack generator to guide the parameter update of the watermark attack generator. Through this feedback mechanism, the watermark attack generator continuously learns how to better imitate the effect of watermark damage, so as to eventually generate damaged watermark information that can deceive the watermark attack discriminator;

[0108] S46: Output of damaged watermark information: After adversarial training and iterative optimization, the watermark attack generator can eventually generate watermark information that is visually similar to the reversed watermark information, so that it can deceive the watermark attack discriminator.

[0109] The entire algorithm flow is a dynamic learning and adaptation process, which requires fine-tuning of the training parameters of the generative adversarial network to ensure that the generated damaged watermark information effectively destroys the integrity and detectability of the watermark.

[0110] S5: Image fusion: The detail-preserving image and the watermark-destroyed image are added pixel-wise to produce the final post-attack image. This step ensures that the visual details of the image are preserved while the watermark information is effectively removed or destroyed.

[0111] The image after the watermark is destroyed AI and detail preservation map Fusion into post-attack image WIAA , which can be achieved by fusion function F To achieve this, this function will destroy the image after the watermark AI and detail-preserving images DPI As input, and output the fused image WIAA , this process can be expressed by the following formula:

[0112] (6)

[0113] in: F is the fusion function;

[0114] are the deep learning model parameters;

[0115] S6: Post-processing and optimization: After obtaining the attacked image, some post-processing steps are required, such as image quality enhancement and noise reduction, to further improve the attack effect and the visual quality of the image.

[0116] The entire algorithm process is an iterative and learning process, which requires a large amount of training data to optimize the parameters of the generative adversarial network to ensure that the final output image not only removes the watermark information but also maintains a high degree of visual fidelity.

[0117] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A dual-branch generation algorithm to counter watermark attacks, characterized in that: The following steps are involved: S1: Construct a sophisticated two-branch GAN architecture, including a detail-preserving GAN and a watermark-attack GAN; S2: Image difference analysis: A sophisticated two-branch generative adversarial network architecture is used to perform pixel-by-pixel difference operations between the original image and the watermarked image to extract image details and watermark information. The output is two independent channels, namely the image detail map and the watermark information map. Watermarked image WI and original watermark information OI Subtract to get image details ; （1） Watermarked image WI and the original image O Subtract and get watermark information I ; （2） S3: Generate detail-preserving image: The image detail map is then input into the detail-preserving generative adversarial network, which generates a detail-preserving image; S4: Generate a watermark-destroyed image: The watermark information graph and the corresponding inverted watermark information graph are input into the watermark attack generative adversarial network to generate a watermark-destroyed image; S5: Image fusion: The detail-preserving image and the watermark-destroyed image are added pixel-wise to produce the final post-attack image.

2. The dual-branch generation anti-watermark attack algorithm according to claim 1 is characterized in that: S3 includes: in the process of generating adversarial, designing a generator G , which maps the image details As input, it generates an image , the image not only retains the features of the detail image, but also should be close to the original image ; At the same time, design a discriminator , distinguish the generated images and the original image ; The following is the objective function of the generator: （3） in: It is a confrontational loss; It is the loss of content; It is the loss of detail; is the target edge enhancement loss; is a weight parameter used to balance the contribution of different loss terms; The following is the objective function of the discriminator: Discriminator The goal is to maximize the following function to correctly distinguish the original image from the generated image: （4） Among them: The first term is the loss of the original image, which encourages the discriminator For the original image O Give high ratings; The second term is the loss of generated images, which encourages the discriminator For the generator G Generated Image Give a low rating.

3. The dual-branch generation anti-watermark attack algorithm according to claim 1 is characterized in that: The S4 includes: the watermark information is I , set the anti-watermark information to RI , the output is the image after watermark destruction AI , the formula for generating adversarial is as follows: （5） Generator G Receive two inputs, the watermark information is I And the anti-watermark information is RI ; Set the anti-watermark information to RI is the image after counter-watermarking, and the generator G Through its parameters The watermark information is I The anti-watermark information is RI Learning and generating watermark-corrupted images AI .

4. The dual-branch generation anti-watermark attack algorithm according to claim 1 is characterized in that: The S5 comprises: AI and detail preservation map Fusion into post-attack image WIAA , which can be achieved by fusion function F To achieve this, this function will destroy the image after the watermark AI and detail-preserving images DPI As input, and output the fused image WIAA , this process can be expressed by the following formula: （6） in: F is the fusion function; are the deep learning model parameters.

5. The dual-branch generation anti-watermark attack algorithm according to claim 1 is characterized in that: The detail-preserving generative adversarial network consists of a detail-preserving generator and a detail-preserving discriminator, which work together to generate high-quality detail-preserving images.

6. The dual-branch generation algorithm for resisting watermark attacks according to claim 2 is characterized in that: The specific process of S3 is as follows: S31: Detail extraction by detail-preserving generator: The detail-preserving generator receives the image detail map as input. The goal of the detail-preserving generator is to use the detail information to generate a detail-preserving image that is visually similar to the original image while removing the watermark information. S32: authenticity judgment of detail-preserving discriminator: the generated detail-preserving image is input to the detail-preserving discriminator together with the original image, and the detail-preserving discriminator provides feedback to guide the detail-preserving generator to optimize the generated detail-preserving image so that it cannot be distinguished from the original image visually; S33: Adversarial training: The detail-preserving generator and detail-preserving discriminator compete with each other in the adversarial training process; S34: Iterative optimization of detail-preserving generator: Based on the feedback from the detail-preserving discriminator, the detail-preserving generator adjusts internal parameters through back-propagation and gradient descent optimization algorithms to generate more realistic detail-preserving images, until the images generated by the detail-preserving generator can effectively remove watermark information without sacrificing visual quality; S35: Result verification: Finally, the generated detail-preserving image is evaluated by the detail-preserving discriminator. If it can frequently deceive the detail-preserving discriminator, that is, it is judged as a real image, then it is considered that the detail-preserving generator has successfully completed the task.

7. The dual-branch generation anti-watermark attack algorithm according to claim 3 is characterized in that: The watermark attack generative adversarial network includes a watermark attack generator and a watermark attack discriminator.

8. The dual-branch generation algorithm for resisting watermark attacks according to claim 3 is characterized in that: The specific process of S4 is as follows: S41: Initialization of watermark attack generator: The watermark attack generator receives the original watermark information as input; S42: Generation of damaged watermark information: The watermark attack generator processes the input original watermark information through its internal network structure to generate damaged watermark information; S43: True or false judgment of watermark attack discriminator: The generated damaged watermark information is then input into the watermark attack discriminator together with the corresponding reversed watermark information. The watermark attack discriminator is responsible for evaluating the two types of information and making a true or false judgment; S44: Adversarial training mechanism: The watermark attack generator and the watermark attack discriminator compete with each other in adversarial training; S45: Feedback optimization of watermark attack generator: the true or false judgment result of the watermark attack discriminator is fed back to the watermark attack generator to guide the parameter update of the watermark attack generator; S46: Output of damaged watermark information: After adversarial training and iterative optimization, the watermark attack generator can eventually generate watermark information that is visually similar to the reversed watermark information, so that it can deceive the watermark attack discriminator.

9. The dual-branch generation anti-watermark attack algorithm according to claim 1 is characterized in that: The reverse watermark information graph is the reverse representation of the watermark.

Images