Watermark anti-counterfeiting evaluation method and system based on diffusion model, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of watermark anti-counterfeiting evaluation, and provides a watermark anti-counterfeiting evaluation method and system based on a diffusion model, electronic equipment and a storage medium, and the method comprises the steps of auxiliary watermark data set generation, image watermark removal, unconditional diffusion model training, watermark information reverse counterfeiting, counterfeited watermark verification and clustering analysis. According to the method, the unconditional diffusion model is used for carrying out watermark imitation on the homologous watermarking-free image, so that adaptive errors caused by manual data selection are avoided, and the learning effect of the unconditional diffusion model on the target object is improved; through a watermark removing model and taking a verification result of a target object as a model iteration loss function, independent processing of suspicious watermark features is realized, and a watermark removing effect is improved; through reverse watermark counterfeiting, counterfeited watermark verification and clustering analysis, the effect of evaluating the watermark counterfeiting resistance of the target object is improved, and the potential watermark anti-counterfeiting problem is visually displayed.

Description

Technical Field

[0001] This invention relates to the field of watermark anti-counterfeiting evaluation technology, and in particular to a watermark anti-counterfeiting evaluation method, system, electronic device and storage medium based on a diffusion model. Background Technology

[0002] With the rapid development of generative artificial intelligence technology, deep learning-based image generation models are widely used in the creation of various visual content. However, with the popularization of the technology, problems such as the difficulty in distinguishing the authenticity of AI-generated content and the difficulty in tracing responsibility have gradually emerged. To ensure the traceability and copyright protection of generated content, image watermarking technology has emerged. Watermarking embeds imperceptible identification information into generated images, enabling the tracking of content sources and copyright verification, and has become an important technical means in the field of AI image security. Currently, image watermarking technology is mainly divided into two categories: one is based on traditional frequency domain or spatial domain transformation methods, such as discrete wavelet transform and discrete cosine transform, which modulate the frequency features of the image to embed watermark information; the other is based on the encoding and decoding structure of deep neural networks, which uses watermark information as network input and is trained to achieve robust and covert watermark embedding and extraction. These methods can resist common image compression, cropping, and noise attacks to a certain extent, improving the covertness and stability of watermarks.

[0003] However, existing invisible watermarking technologies still face serious security vulnerabilities in practical applications, particularly in watermark forgery attack scenarios. Existing forgeries typically rely on partial or full access to the watermarking algorithm, such as black-box queries or white-box gradient information, lacking effective attack methods for unboxed environments where access is restricted. This lack of effective attack methods leads to insufficient evaluation of the anti-counterfeiting capabilities of watermarking systems, potentially affecting the credibility of watermark-based AI content responsibility attribution and copyright protection mechanisms. Furthermore, current watermark anti-counterfeiting evaluations lack a unified and systematic evaluation framework; most studies are conducted only in single watermarking schemes or idealized non-unboxed environments, failing to comprehensively reflect security risks in real-world application environments. Moreover, existing verification methods often use watermark-free data unrelated to the verification object, but different verification objects have varying adaptability to different images, and human selection can reduce the reliability of evaluation results. In addition, conventional methods only perform simple verification of the verification object, lacking corresponding analytical processes regarding the relationship between the verification image and the result, and how to improve it. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a watermark anti-counterfeiting evaluation method, system, electronic device, and storage medium based on a diffusion model, thereby solving the problem of the lack of evaluation of watermark forgery attack scenarios in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A watermark anti-counterfeiting evaluation method based on a diffusion model includes:

[0007] Generative AI images of the target object containing hidden watermarks are collected to obtain an auxiliary watermark dataset;

[0008] The watermark is removed from the images in the auxiliary watermark dataset using a pre-built watermark removal model to obtain a watermark-free dataset.

[0009] The pre-built unconditional diffusion model is trained using the auxiliary watermark dataset based on the DDIM framework.

[0010] The trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information.

[0011] The forged image set is input into the target object for verification, and the verification result is obtained;

[0012] The verification results are analyzed using a clustering analysis model based on the preset label data of the auxiliary watermark dataset to obtain a watermark anti-counterfeiting evaluation report.

[0013] Preferably, generative AI images of the target object containing invisible watermarks are collected to obtain an auxiliary watermark dataset, including:

[0014] Define the data collection time range, target domain, and target website to obtain the data crawling requirements;

[0015] Data to be analyzed is collected according to the data crawling requirements, and the data to be analyzed is filtered by images to obtain image information; the image information includes: image number and image title;

[0016] By filtering the text information corresponding to the image information in the data to be analyzed, the text to be analyzed is obtained;

[0017] The text to be analyzed is cleaned to obtain denoised text;

[0018] The TF-IDF formula was used to perform word frequency statistics on the denoised text to obtain word frequency data;

[0019] Based on the word frequency data, keywords whose word frequencies meet preset requirements are extracted from the denoised text to obtain a prompt word dataset;

[0020] The prompt word dataset is input into the target object to generate an image and an invisible watermark, thus obtaining the auxiliary watermark dataset.

[0021] Preferably, the process of constructing the watermark removal model includes:

[0022] Construct an original model; the original model includes: a dynamic feature decoupling encoder and a hierarchical decoder connected in sequence; the dynamic feature decoupling encoder includes: a multi-scale backbone extraction layer and a feature decoupling discriminator connected in sequence; the hierarchical decoder includes: a fusion module and pixel-level branches, texture-level branches, and semantic-level branches respectively connected to the fusion module;

[0023] The multi-scale backbone extraction layer is used to extract full-dimensional features from the auxiliary watermark dataset to obtain the original extracted features.

[0024] The feature decoupling discriminator is used to split the original extracted features into a content feature stream and a suspicious watermark feature stream;

[0025] The suspicious watermark feature stream is modified by performing abnormal pixel modification and abnormal texture modification using the pixel-level branch and the texture-level branch respectively to obtain the modified feature stream. The content feature stream and the modified feature stream are then decoded using the pixel-level branch, the texture-level branch and the semantic-level branch respectively to obtain pixel decoded data, texture decoded data and semantic decoded data.

[0026] The fusion module is used to fuse the pixel decoding data, texture decoding data, and semantic decoding data to obtain a reconstructed image.

[0027] Based on the reconstructed image and the dewatermarked dataset, the original model is iterated using a dewatermarking loss function to obtain the dewatermarked model; the expression for the dewatermarking loss function is: ;in, This is the output of the watermark removal loss function; To reconstruct the similarity loss; The watermark detection loss is set to 0 when the detection result of inputting the reconstructed image into the target object is that no watermark is detected, and 1 otherwise.

[0028] Preferably, training a pre-built unconditional diffusion model using the auxiliary watermark dataset based on the DDIM framework includes:

[0029] The noise prediction function of the auxiliary watermark dataset is learned using the unconditional diffusion model, and the neural network parameters of the unconditional diffusion model are optimized based on the DDIM framework.

[0030] Preferably, the trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, including:

[0031] By utilizing the noise prediction capability of the trained unconditional diffusion model, the feature signal of the target watermark is injected into the potential representation by adjusting the sampling path in the noise space, thereby obtaining the forged image set containing the watermark information.

[0032] Preferably, the trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, further comprising:

[0033] The generated set of forged images is iteratively optimized using a diffusion model scoring function, and artifacts in the target watermark injection process of the forged image set are eliminated by maximizing the probability of the image on the target watermark distribution and the visual similarity between the image and the original input.

[0034] Preferably, the verification results are analyzed using a clustering analysis model based on the preset label data of the auxiliary watermark dataset to obtain a watermark anti-counterfeiting evaluation report, including:

[0035] By integrating the preset label data and the verification results, an analysis sample set is obtained;

[0036] Based on the analyzed sample set, construct watermark anti-counterfeiting features, generate correlation features, and generate parameter correlation features;

[0037] Based on the preset improvement scheme library, cluster analysis model is used to perform cluster analysis on the watermark anti-counterfeiting features, the generation correlation features, and the generation parameter correlation features to obtain the overall anti-counterfeiting performance index, the high-risk type list, and the candidate improvement schemes corresponding to the high-risk type list;

[0038] By integrating the overall anti-counterfeiting performance indicators, the list of high-risk types, and the candidate improvement schemes, a watermark anti-counterfeiting evaluation report is obtained.

[0039] Preferably, a watermark anti-counterfeiting evaluation system based on a diffusion model includes:

[0040] The image collection module is used to collect generative AI images of the target object containing invisible watermarks to obtain an auxiliary watermark dataset.

[0041] The watermark-free homologous image generation module is used to remove watermarks from images in the auxiliary watermark dataset using a pre-built watermark removal model to obtain a watermark-free dataset.

[0042] The model training module is used to train a pre-built unconditional diffusion model based on the DDIM framework using the auxiliary watermark dataset.

[0043] The boxless environment watermark simulation module is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset using the trained unconditional diffusion model to obtain a set of forged images containing watermark information.

[0044] The anti-counterfeiting performance verification module is used to input the set of counterfeit images into the target object for verification and obtain the verification result.

[0045] The result evaluation module is used to analyze the verification results based on the preset label data of the auxiliary watermark dataset using a clustering analysis model, and obtain a watermark anti-counterfeiting evaluation report.

[0046] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned diffusion-based watermark anti-counterfeiting evaluation method.

[0047] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned diffusion-based watermark anti-counterfeiting evaluation method.

[0048] The present invention discloses the following technical effects:

[0049] This invention provides a watermark anti-counterfeiting evaluation method, system, electronic device, and storage medium based on a diffusion model. By using an unconditional diffusion model to forge watermarks on watermark-free images from the same source, it solves the problem of adaptive errors caused by manually selected data and improves the learning effect on the target object. By using a watermark removal model and taking the verification result of the target object as the model iteration loss function, it solves the problem of lacking evaluation of watermark forgery attack scenarios in the evaluation process and realizes the separate processing of suspicious watermark features. Through reverse watermark forgery, forged watermark verification, and cluster analysis, it solves the problem of lacking corresponding analysis processes on the relationship between verification images and results and how to improve them, and realizes the extraction of potential watermark anti-counterfeiting problems. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the watermark anti-counterfeiting evaluation process based on the diffusion model provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the auxiliary watermark dataset generation process provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the watermark removal model construction process provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the clustering analysis process provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of a watermark anti-counterfeiting evaluation system based on a diffusion model, provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The purpose of this invention is to provide a watermark anti-counterfeiting evaluation method, system, electronic device and storage medium based on a diffusion model, to solve the problem of the lack of evaluation of watermark forgery attack scenarios in the existing technology.

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] Figure 1 This is a schematic diagram of the watermark anti-counterfeiting evaluation process based on the diffusion model provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a watermark anti-counterfeiting evaluation method based on a diffusion model, comprising:

[0060] Step 100: Collect generative AI images of the target object containing invisible watermarks to obtain an auxiliary watermark dataset;

[0061] Step 200: Use a pre-built watermark removal model to remove watermarks from the images in the auxiliary watermark dataset to obtain a watermark-removed dataset;

[0062] Step 300: Train the pre-built unconditional diffusion model using the auxiliary watermark dataset based on the DDIM framework;

[0063] Step 400: Use the trained unconditional diffusion model to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information;

[0064] Step 500: Input the forged image set into the target object for verification, and obtain the verification result;

[0065] Step 600: Analyze the verification results using a clustering analysis model based on the preset label data of the auxiliary watermark dataset to obtain a watermark anti-counterfeiting evaluation report.

[0066] refer to Figure 2 Generative AI images of the target object containing hidden watermarks are collected to obtain an auxiliary watermark dataset, including:

[0067] Step 101: Set the collection time range, target field, and target website to obtain the data crawling requirements;

[0068] Step 102: Collect the data to be analyzed according to the data crawling requirements, and perform image filtering on the data to be analyzed to obtain image information; the image information includes: image number and image title;

[0069] Step 103: Filter the text information in the data to be analyzed that corresponds to the image information to obtain the text to be analyzed;

[0070] Step 104: Perform data cleaning on the text to be analyzed to obtain denoised text;

[0071] Step 105: Perform word frequency statistics on the denoised text using the TF-IDF formula to obtain word frequency data;

[0072] Step 106: Extract keywords whose word frequencies meet preset requirements from the denoised text based on the word frequency data to obtain a prompt word dataset;

[0073] Step 107: Input the prompt word dataset into the target object to generate an image and generate an invisible watermark, thereby obtaining the auxiliary watermark dataset.

[0074] refer to Figure 3 The process of constructing the watermark removal model includes:

[0075] Step 201: Construct the original model; the original model includes: a dynamic feature decoupling encoder and a hierarchical decoder connected in sequence; the dynamic feature decoupling encoder includes: a multi-scale backbone extraction layer and a feature decoupling discriminator connected in sequence; the hierarchical decoder includes: a fusion module and pixel-level branches, texture-level branches, and semantic-level branches connected to the fusion module respectively;

[0076] Step 202: Use the multi-scale backbone extraction layer to extract full-dimensional features from the auxiliary watermark dataset to obtain the original extracted features;

[0077] Step 203: Use the feature decoupling discriminator to split the original extracted features into a content feature stream and a suspicious watermark feature stream;

[0078] Step 204: Modify the suspicious watermark feature stream by performing abnormal pixel modification and abnormal texture modification using the pixel-level branch and the texture-level branch respectively to obtain the modified feature stream, and decode the content feature stream and the modified feature stream using the pixel-level branch, the texture-level branch and the semantic-level branch respectively to obtain pixel decoded data, texture decoded data and semantic decoded data;

[0079] Step 205: Use the fusion module to fuse the pixel decoding data, texture decoding data, and semantic decoding data to obtain the reconstructed image;

[0080] Step 206: Based on the reconstructed image and the dewatermarked dataset, iterate the original model using the dewatermarking loss function to obtain the dewatermarked model; the expression for the dewatermarking loss function is: ;in, This is the output of the watermark removal loss function; To reconstruct the similarity loss; The watermark detection loss is set to 0 when the detection result of inputting the reconstructed image into the target object is that no watermark is detected, and 1 otherwise.

[0081] Specifically, the pre-built unconditional diffusion model is trained using the auxiliary watermark dataset based on the DDIM framework, including:

[0082] The noise prediction function of the auxiliary watermark dataset is learned using the unconditional diffusion model, and the neural network parameters of the unconditional diffusion model are optimized based on the DDIM framework.

[0083] Furthermore, the trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, including:

[0084] By utilizing the noise prediction capability of the trained unconditional diffusion model, the feature signal of the target watermark is injected into the potential representation by adjusting the sampling path in the noise space, thereby obtaining the forged image set containing the watermark information.

[0085] Specifically, the trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, and the method further includes:

[0086] The generated set of forged images is iteratively optimized using a diffusion model scoring function, and artifacts in the target watermark injection process of the forged image set are eliminated by maximizing the probability of the image on the target watermark distribution and the visual similarity between the image and the original input.

[0087] refer to Figure 4 The verification results are analyzed using a clustering analysis model based on the preset label data of the auxiliary watermark dataset to obtain a watermark anti-counterfeiting evaluation report, including:

[0088] Step 601: Integrate the preset label data and the verification results to obtain the analysis sample set;

[0089] Step 602: Construct watermark anti-counterfeiting features, generate correlation features, and generate parameter correlation features based on the analyzed sample set;

[0090] Step 603: Based on the preset improvement scheme library, use a cluster analysis model to perform cluster analysis on the watermark anti-counterfeiting features, the generation correlation features, and the generation parameter correlation features to obtain the overall anti-counterfeiting performance index, the high-risk type list, and the candidate improvement schemes corresponding to the high-risk type list;

[0091] Step 604: Integrate the overall anti-counterfeiting performance indicators, the high-risk type list, and the candidate improvement schemes to obtain a watermark anti-counterfeiting evaluation report.

[0092] Specifically, the time frame, target domain, and websites for data collection are determined, and the requirements for data scraping are clarified. The time frame needs to include updates from the target object over the past six months, and the target domain can be further subdivided into technology, education, entertainment, etc. Data to be analyzed is collected according to the data scraping requirements. Then, images are filtered to obtain image information. Image filtering is performed by judging the file format. In this embodiment, it verifies whether the file extension is in a preset image format (including .jpg, .png, .webp, .bmp, etc.), checks whether the file header identifier and extension are consistent, excludes fake files, and removes non-image data that fails format verification. Image information includes image number, title, original file format, and storage path. From the data to be analyzed, text information related to the image information is filtered out to obtain the text to be analyzed. Text information includes image descriptions, tags filled in by the user during upload, and titles and summaries of relevant pages. This text is then cleaned to obtain noise-free text. Data cleaning includes deleting duplicate text entries, removing HTML tags and special symbols, filtering out meaningless function words, and correcting text encoding errors. The TF-IDF formula was used to perform word frequency statistics on the cleaned text to obtain word frequency data. During word frequency statistics, words in the stop word list were excluded; the stop word list contains common, non-discriminatory words in the target domain. Based on the word frequency data, keywords that ranked in the top 30% of the cleaned text and appeared at least 5 times were extracted. Keywords with similar meanings were merged to obtain a prompt word dataset. The prompt word dataset was divided into groups and input into the target object in batches. Uniform image generation parameters were set to trigger the target object's image generation function and built-in invisible watermark embedding function, ultimately resulting in an auxiliary watermark dataset.

[0093] Preferably, this embodiment designs a watermark removal model, which consists of the following parts: a dynamic feature decoupling encoder and a hierarchical decoder. The dynamic feature decoupling encoder includes a multi-scale backbone extraction layer and a feature decoupling discriminator. The hierarchical decoder contains a fusion module and three branches: pixel-level, texture-level, and semantic-level, all of which are connected to the fusion module. The multi-scale backbone extraction layer uses an improved ResNet-50 network, extracting feature maps at 1 / 2, 1 / 4, 1 / 8, and 1 / 16 resolutions through four stages of convolutional operations. This covers both local details of the image and takes into account global structural information. The feature decoupling discriminator includes parallel content feature discrimination subnetworks and watermark feature discrimination subnetworks, which separate the feature streams through adversarial training.

[0094] Furthermore, this embodiment uses a multi-scale backbone extraction layer to extract features from the auxiliary watermark dataset, extracting the original features. These features include pixel value distribution, texture gradient, semantic region contours, and high-frequency noise features, etc. By stitching together the multi-scale feature maps, cross-dimensional information fusion is achieved. Then, a feature decoupling discriminator is used to divide the original extracted features into a content feature stream and a suspicious watermark feature stream. This process is achieved through contrastive learning: the content feature discriminator subnetwork outputs the probability that the feature belongs to the image content, and the watermark feature discriminator subnetwork outputs the probability that the feature belongs to the watermark. By minimizing the mutual information entropy between the content features and the watermark features, these two types of feature streams are separated in high-dimensional space. The pixel-level branch and the texture-level branch modify abnormal pixels and correct abnormal textures in the suspicious watermark feature stream, respectively, to obtain the modified feature stream. Specifically, the pixel-level branch locates abnormal pixels related to the watermark through 1×1 convolution and skip connections, and then replaces these pixels with neighborhood pixel weighted interpolation. The texture-level branch captures abnormal texture regions through a non-local attention module, and then uses a generative adversarial network to generate corrected textures that conform to the surrounding texture distribution. Finally, the pixel-level branch, texture-level branch, and semantic-level branch decode the content feature stream and the modified feature stream, respectively, to obtain pixel-decoded data, texture-decoded data, and semantic-decoded data. The pixel-level branch uses a transposed convolutional layer to achieve 2x upsampling, outputting detail data with the same resolution as the original image; the texture-level branch outputs local texture consistency features; and the semantic-level branch uses global average pooling and class activation mapping (CAM) to locate core semantic regions, outputting semantically preserved structural data.

[0095] Optionally, a fusion module is used to fuse pixel-level decoded data, texture-level decoded data, and semantic-level decoded data to obtain a reconstructed image. The fusion module employs a dynamic weight allocation mechanism: weight coefficients are calculated based on the gradient values ​​of pixel-level data, the entropy values ​​of texture-level data, and the confidence levels of semantic-level data, and then weighted summation is used to achieve adaptive fusion of multi-dimensional features. Then, based on the reconstructed image and the dewatermarked dataset, the original model is iterated using a dewatermarking loss function to obtain the final dewatermarking model.

[0096] Specifically, an unconditional diffusion model is trained using an auxiliary dataset, enabling the model to learn the noise prediction function of the target watermark image, thereby implicitly representing the distribution characteristics of the watermark information in the image space. During training, the DDIM framework is employed, and high-precision fitting of the watermark distribution is achieved by optimizing the neural network parameters. For the unwatermarked image to be tested, a shallow DDIM inverse process is performed, mapping the image to the latent noise space as an intermediate step. This avoids semantic information loss caused by deep inverse processing, preserving the original visual information of the image and ensuring the quality foundation for subsequent image evaluation. Combining the trained diffusion model, its noise prediction capability guides the inverse process. By adjusting the sampling path in the noise space, the feature signal of the target watermark is injected into the latent representation, generating a forged image containing watermark information. Iterative optimization based on the diffusion model's scoring function is applied to the generated forged watermark image, jointly maximizing the probability of the image on the target watermark distribution and the visual similarity between the image and the original input. This further eliminates artifacts during the target watermark injection process, improving the visual quality of the forged watermark image and the watermark detection effect. The method described above was used to verify the generality and effectiveness of various mainstream invisible watermarking schemes (open source and closed source).

[0097] Preferably, to construct the analysis sample set, pre-defined label data is processed, including image generation parameters, domain labels, and prompt word association labels. Verification results include the detection pass rate, false positive rate, and false negative rate of the counterfeit watermark. Then, the pre-defined label data and verification results are associated according to image sequence number to form a structured analysis sample set. Based on the verification results, the anti-counterfeiting capability of the watermark is calculated. Generation correlation is obtained through cross-analysis of the pre-defined label data and verification results. Quantifying the impact of generation parameters on anti-counterfeiting performance yields the generation parameter association features. A high-risk type determination rule is defined: if the counterfeiting success rate of samples in a cluster exceeds 70%, it is marked as a high-risk cluster. Common features of these high-risk clusters are extracted to form a high-risk type list. Finally, an overall anti-counterfeiting performance radar chart, a heatmap of anti-counterfeiting performance in each domain, and a pie chart of high-risk type distribution are output. A decision tree model is used to establish the correspondence between high-risk features and improvement schemes, outputting alternative schemes for each high-risk type and displaying them in order of expected effect.

[0098] refer to Figure 5 As an optional implementation, this embodiment also provides a watermark anti-counterfeiting evaluation system based on a diffusion model, including:

[0099] The image collection module is used to collect generative AI images of the target object containing invisible watermarks to obtain an auxiliary watermark dataset.

[0100] The watermark-free homologous image generation module is used to remove watermarks from images in the auxiliary watermark dataset using a pre-built watermark removal model to obtain a watermark-free dataset.

[0101] The model training module is used to train a pre-built unconditional diffusion model based on the DDIM framework using the auxiliary watermark dataset.

[0102] The boxless environment watermark simulation module is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset using the trained unconditional diffusion model to obtain a set of forged images containing watermark information.

[0103] The anti-counterfeiting performance verification module is used to input the set of counterfeit images into the target object for verification and obtain the verification result.

[0104] The result evaluation module is used to analyze the verification results based on the preset label data of the auxiliary watermark dataset using a clustering analysis model, and obtain a watermark anti-counterfeiting evaluation report.

[0105] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned diffusion model-based watermark anti-counterfeiting evaluation method.

[0106] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned diffusion-based watermark anti-counterfeiting evaluation method.

[0107] The beneficial effects of this invention are as follows:

[0108] This invention uses an unconditional diffusion model to replicate watermarks on homologous watermark-free images, avoiding adaptive errors caused by manually selected data and improving the learning effect of the unconditional diffusion model on the target object. By using a watermark removal model and taking the verification result of the target object as the model iteration loss function, it achieves separate processing of suspicious watermark features, improving the watermark removal effect. Through reverse watermark forgery, forged watermark verification, and cluster analysis, it improves the evaluation effect of the anti-watermark forgery performance of the target object and intuitively demonstrates the potential watermark anti-counterfeiting problem.

[0109] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0110] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A watermark anti-counterfeiting evaluation method based on a diffusion model, characterized in that, include: Generative AI images of the target object containing hidden watermarks are collected to obtain an auxiliary watermark dataset; The watermark is removed from the images in the auxiliary watermark dataset using a pre-built watermark removal model to obtain a watermark-free dataset. The pre-built unconditional diffusion model is trained using the auxiliary watermark dataset based on the DDIM framework. The trained unconditional diffusion model is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset to obtain a set of forged images containing watermark information. The forged image set is input into the target object for verification, and the verification result is obtained; The verification results are analyzed using a clustering analysis model based on the preset label data of the auxiliary watermark dataset to obtain a watermark anti-counterfeiting evaluation report.

2. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 1, characterized in that, Generative AI images of the target object containing hidden watermarks are collected to obtain an auxiliary watermark dataset, including: Define the data collection time range, target domain, and target website to obtain the data crawling requirements; Data to be analyzed is collected according to the data crawling requirements, and the data to be analyzed is filtered by images to obtain image information; the image information includes: image number and image title; By filtering the text information corresponding to the image information in the data to be analyzed, the text to be analyzed is obtained; The text to be analyzed is cleaned to obtain denoised text; The TF-IDF formula was used to perform word frequency statistics on the denoised text to obtain word frequency data; Based on the word frequency data, keywords whose word frequencies meet preset requirements are extracted from the denoised text to obtain a prompt word dataset; The prompt word dataset is input into the target object to generate an image and an invisible watermark, thus obtaining the auxiliary watermark dataset.

3. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 2, characterized in that, The process of constructing the watermark removal model includes: Construct an original model; the original model includes: a dynamic feature decoupling encoder and a hierarchical decoder connected in sequence; the dynamic feature decoupling encoder includes: a multi-scale backbone extraction layer and a feature decoupling discriminator connected in sequence; the hierarchical decoder includes: a fusion module and pixel-level branches, texture-level branches, and semantic-level branches respectively connected to the fusion module; The multi-scale backbone extraction layer is used to extract full-dimensional features from the auxiliary watermark dataset to obtain the original extracted features. The feature decoupling discriminator is used to split the original extracted features into a content feature stream and a suspicious watermark feature stream; The suspicious watermark feature stream is modified by performing abnormal pixel modification and abnormal texture modification using the pixel-level branch and the texture-level branch respectively to obtain the modified feature stream. The content feature stream and the modified feature stream are then decoded using the pixel-level branch, the texture-level branch and the semantic-level branch respectively to obtain pixel decoded data, texture decoded data and semantic decoded data. The fusion module is used to fuse the pixel decoding data, texture decoding data, and semantic decoding data to obtain a reconstructed image. Based on the reconstructed image and the dewatermarked dataset, the original model is iterated using a dewatermarking loss function to obtain the dewatermarked model; the expression for the dewatermarking loss function is: ;in, This is the output of the watermark removal loss function; To reconstruct the similarity loss; The watermark detection loss is set to 0 when the detection result of inputting the reconstructed image into the target object is that no watermark is detected, and 1 otherwise.

4. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 3, characterized in that, Based on the DDIM framework, the pre-built unconditional diffusion model is trained using the aforementioned auxiliary watermark dataset, including: The noise prediction function of the auxiliary watermark dataset is learned using the unconditional diffusion model, and the neural network parameters of the unconditional diffusion model are optimized based on the DDIM framework.

5. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 4, characterized in that, Using the trained unconditional diffusion model, shallow DDIM inverse processing is performed on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, including: By utilizing the noise prediction capability of the trained unconditional diffusion model, the feature signal of the target watermark is injected into the potential representation by adjusting the sampling path in the noise space, thereby obtaining the forged image set containing the watermark information.

6. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 5, characterized in that, Using the trained unconditional diffusion model, shallow DDIM inverse processing is performed on the images in the watermark removal dataset to obtain a set of forged images containing watermark information, which also includes: The generated set of forged images is iteratively optimized using a diffusion model scoring function, and artifacts in the target watermark injection process of the forged image set are eliminated by maximizing the probability of the image on the target watermark distribution and the visual similarity between the image and the original input.

7. The watermark anti-counterfeiting evaluation method based on a diffusion model according to claim 6, characterized in that, Based on the preset label data of the auxiliary watermark dataset, the verification results are analyzed using a clustering analysis model to obtain a watermark anti-counterfeiting evaluation report, including: By integrating the preset label data and the verification results, an analysis sample set is obtained; Based on the analyzed sample set, construct watermark anti-counterfeiting features, generate correlation features, and generate parameter correlation features; Based on the preset improvement scheme library, cluster analysis model is used to perform cluster analysis on the watermark anti-counterfeiting features, the generation correlation features, and the generation parameter correlation features to obtain the overall anti-counterfeiting performance index, the high-risk type list, and the candidate improvement schemes corresponding to the high-risk type list; By integrating the overall anti-counterfeiting performance indicators, the list of high-risk types, and the candidate improvement schemes, a watermark anti-counterfeiting evaluation report is obtained.

8. A watermark anti-counterfeiting evaluation system based on a diffusion model, characterized in that, include: The image collection module is used to collect generative AI images of the target object containing invisible watermarks to obtain an auxiliary watermark dataset. The watermark-free homologous image generation module is used to remove watermarks from images in the auxiliary watermark dataset using a pre-built watermark removal model to obtain a watermark-free dataset. The model training module is used to train a pre-built unconditional diffusion model based on the DDIM framework using the auxiliary watermark dataset. The boxless environment watermark simulation module is used to perform shallow DDIM inverse processing on the images in the watermark removal dataset using the trained unconditional diffusion model to obtain a set of forged images containing watermark information. The anti-counterfeiting performance verification module is used to input the set of counterfeit images into the target object for verification and obtain the verification result. The result evaluation module is used to analyze the verification results based on the preset label data of the auxiliary watermark dataset using a clustering analysis model, and obtain a watermark anti-counterfeiting evaluation report.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a diffusion-based watermark anti-counterfeiting evaluation method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute a watermark anti-counterfeiting evaluation method based on a diffusion model, as described in any one of claims 1 to 7.

Images