A fingerprint protection method for image denoising based on diffusion probability model

Abstract

The present invention discloses a fingerprint protection method for an image denoising diffusion probability model, belonging to the field of electronic digital data processing technology, comprising: obtaining a protection target verification image; inversely obtaining the latent code of the denoising diffusion probability model corresponding to the protection target verification image #imgabs0# as the fingerprint code of the protection target verification image; obtaining a suspected theft denoising diffusion probability model, inputting the fingerprint code of the protection target verification image into the sampling process of the suspected theft denoising diffusion probability model, and obtaining an output image corresponding to the suspected theft denoising diffusion probability model; and comparing the copyright information of the output image corresponding to the suspected theft denoising diffusion probability model with the copyright information of the protection target verification image. The fingerprint protection method for an image denoising diffusion probability model solves the problem that existing DDPM protection schemes modify DDPM parameters, thereby affecting the performance and performance of DDPM.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a fingerprint protection method for an image denoising diffusion probability model. Background Art

[0002] AIGC stands for Artificial Intelligence Generated Content, which refers to AI-generated content. Its key model, DDPM, stands for Denoising Diffusion Probabilistic Model. It uses an iterative denoising process to map pure noise onto a target distribution for generation. The training process of DDPM for image generation gradually adds noise to a clear image, reducing it to pure noise. A denoiser is trained to estimate the amount of noise added each time, and during generation, the pure noise is denoised in reverse order, step by step, to produce a clear image.

[0003] With the widespread adoption of AIGC, its key model, DDPM, has become very active in the open source community, with model sharing becoming commonplace. Protecting its intellectual property is crucial. Currently, there are cases of DDPM theft in the industry, such as obtaining DDPM from open source projects and illegally using it for commercial purposes through attacks like compression or parameter fine-tuning. Existing protection solutions primarily rely on watermarking schemes, which require modifying DDPM parameters. These schemes, such as embedding backdoors into watermark mappings, are often invasive and can affect the model's performance. Summary of the Invention

[0004] In order to overcome the defects of the prior art, the present invention provides a fingerprint protection method for an image denoising diffusion probability model to solve the above problems.

[0005] The present invention solves the technical problem by adopting a technical solution: a fingerprint protection method for an image denoising diffusion probability model, comprising the following steps:

[0006] S1: Obtaining protection target verification image ; Wherein the protection target verification image Built-in copyright information of the model owner;

[0007] S2: Fixed protection target verification image and protection target verification image Corresponding denoising diffusion probability model , reverse protection target verification image Corresponding denoising diffusion probability model Hidden code Verify image as protection target Fingerprint code;

[0008] S3: Obtaining a denoising diffusion probability model for suspected theft , the protection target verification image The fingerprint encoding is input into the denoising diffusion probability model of suspected theft The sampling process of the suspected theft is used to obtain the denoising diffusion probability model The corresponding output image;

[0009] S4: Comparing the suspected theft denoising diffusion probability model The copyright information of the corresponding output image is verified with the protection target image Copyright information to determine the suspected theft of the denoising diffusion probability model The authenticity of the theft; wherein, when the denoising diffusion probability model of the suspected theft The copyright information of the corresponding output image is verified with the protection target image The denoising diffusion probability model indicates suspected theft when the copyright information is consistent The authenticity of the theft is true, otherwise it means the denoising diffusion probability model of the suspected theft The authenticity of the misappropriation is false.

[0010] It is worth noting that in step S2, the protection target verification image is first Gradually add noise to obtain a set of verification images across the protection target Corresponding denoising diffusion probability model The noisy image of the denoiser performance boundary is used as the boundary crossing trajectory; and the image is verified based on the protection target. Corresponding denoising diffusion probability model The sampling process uses model inversion to obtain the latent coding component.

[0011] Preferably, in step S2, the protection target verification image As a denoising diffusion probability model Step 0 image , for the image Gradually add a set of noise sequences , and obtain a set of noisy image sequences , where T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time;

[0012] By optimizing the noise sequence , so that the noisy image As t increases, it gradually crosses the protection target verification image Corresponding denoising diffusion probability model Denoiser The performance boundary area, and the trajectory image sequence at this time The protection target verification image Corresponding denoising diffusion probability model Denoiser boundary crossing trajectories.

[0013] Optionally, in step S2, the protection target verification image Corresponding denoising diffusion probability model Denoiser , first give the noise at each diffusion time step t Initialize to get , get the initialized noise sequence ;

[0014] ;in, is a sample from a standard Gaussian distribution, is a noise image outside the Gaussian distribution, is the preset weight;

[0015] At each diffusion time step t, the protection target verification image and the denoising diffusion probability model corresponding to the protection target verification image are fixed Denoiser ,by As noise Initialization to optimize the noise , for the image Adding Noise After optimizing the noise, we get the image ;Finally, all the images corresponding to the diffusion time step t Composed set to obtain the verification image crossing the protection target Corresponding denoising diffusion probability model Denoiser Performance boundary region trajectory image sequence .

[0016] Specifically, in step S2, the noise is optimized according to the gradient descent , λ represents the learning rate used in the optimization process; based on gradient descent optimization noise The loss function ;

[0017] in, are the preset weights, operators Refers to the calculation image The gradient of T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time, t is the current diffusion time step, Denoiser Predicting noisy images The noise applied in to predict the optimized noise , Representation norm distance, Representation norm The distance, norm For all elements within the norm The sum of the absolute values, norm is the Euclidean distance.

[0018] It is worth noting that in step S2, the image By the current noise and images Denoised diffusion probability model The diffusion process yields: ,in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

[0019] Optionally, in step S2, the boundary crossing trajectory is used , through the denoising diffusion probability model The generation process of the noise perturbation added at each time step t is reversed , using pure noise and noise disturbance Composition Collection As a hidden code ;

[0020] ;in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in

[0021] The mean estimator is calculated as ; ; , , is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

[0022] Specifically, in step S3, the hidden code Input to the denoised diffusion probability model of suspected theft The generation process , the image at t=1 As output image;

[0023] in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in ; when t=T, Hidden code in ;

[0024] Mean Estimator ; ; , ; is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

[0025] The beneficial effect of the present invention is that in the fingerprint protection method for the image denoising diffusion probability model, the image is verified by using the protection target The fingerprint code is input / injected into the denoising diffusion probability model of suspected theft In the iterative sampling of , the denoising diffusion probability model of suspected theft is observed The fingerprint protection method for the image denoising diffusion probability model is a non-invasive model protection scheme that does not change the protection target verification image at all. Corresponding denoising diffusion probability model any model parameters of , and therefore will not be validated against the protection target image Corresponding denoising diffusion probability model The performance and generation quality of the image restoration model are not affected in any way; it fully utilizes the training, progressive denoising and sampling characteristics of DDPM iterative diffusion; during the verification stage through steps S3 and S4, there is no need to calculate the model gradient, and only black box permission requirements are required for the denoiser, which greatly reduces the white box permission requirements for model parameters and gradients required for the fingerprint scheme of the image restoration model in the existing technology; the fingerprint protection method for the image denoising diffusion probability model supports customized images in the verification process, and the verification process is very intuitive and convenient. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a fingerprint protection method for an image denoising diffusion probability model in one embodiment of the present invention;

[0027] Figure 2 This is an overall fingerprint protection framework for a fingerprint protection method for an image denoising diffusion probability model in one embodiment of the present invention.

[0028] Figure 3 Schematic diagram of an abstract definition of a boundary crossing trajectory in one embodiment of the present invention;

[0029] Figure 4 A schematic diagram of a fingerprint extraction process in one embodiment of the present invention;

[0030] Figure 5 Schematic diagram of a fingerprint verification process in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0032] like Figure 1-5 As shown, a fingerprint protection method for an image denoising diffusion probability model includes the following steps:

[0033] S1: Obtaining protection target verification image ; Wherein the protection target verification image The copyright information of the model owner is built in, such as a QR code image;

[0034] S2: If Figure 4 As shown, the fixed protection target verification image and protection target verification image Corresponding denoising diffusion probability model , so that the protection target verification image and protection target verification image Corresponding denoising diffusion probability model Unchanged, reverse protection target verification image Corresponding denoising diffusion probability model Hidden code Verify image as protection target Fingerprint code;

[0035] S3: If Figure 5 As shown, the denoised diffusion probability model of suspected theft is obtained , the protection target verification image The fingerprint encoding is input into the denoising diffusion probability model of suspected theft The sampling process of the suspected theft is used to obtain the denoising diffusion probability model The corresponding output image;

[0036] S4: Comparing the suspected theft denoising diffusion probability model The copyright information of the corresponding output image is verified with the protection target image Copyright information to determine the suspected theft of the denoising diffusion probability model The authenticity of the theft; wherein, when the denoising diffusion probability model of the suspected theft The copyright information of the corresponding output image is verified with the protection target image The denoising diffusion probability model indicates suspected theft when the copyright information is consistent The authenticity of the theft is true, otherwise it means the denoising diffusion probability model of the suspected theft The authenticity of the theft is false. In judging the denoising diffusion probability model of suspected theft When verifying the authenticity of the theft, the denoising diffusion probability model of suspected theft can be used The copyright information of the corresponding output image and the protection target verification image The consistency of the copyright information of the two can be judged, such as by scanning the QR code images output by the two, or performing a statistical test on the similarity between the two.

[0037] In the fingerprint protection method for the image denoising diffusion probability model, the image is verified by using the protection target The fingerprint code is input / injected into the denoising diffusion probability model of suspected theft In the iterative sampling of , the denoising diffusion probability model of suspected theft is observed Therefore, in this solution, we can define the protection target verification image To include the denoising diffusion probability model The QR code image of the owner's copyright information is used to verify the denoising diffusion probability model for suspicious theft. The output image only needs to be scanned to see if the scan result contains the copyright information of the owner. If it does, then there is theft, otherwise there is no such thing. This solution can also verify the image by calculating the protection target. Denoising Diffusion Probability Model for Suspected Theft The similarity of the output image is then determined by statistical testing to determine whether there is any theft. If the protection target verification image Denoising Diffusion Probability Model for Suspected Theft If the output image is too similar, there is theft, otherwise there is no theft. The fingerprint protection method for the image denoising diffusion probability model is a non-invasive model protection scheme that does not change the protection target verification image at all. Corresponding denoising diffusion probability model any model parameters of , and therefore will not be validated against the protection target image Corresponding denoising diffusion probability model The performance and generation quality of the image denoising diffusion model are not affected in any way; it makes full use of the training, progressive denoising and sampling characteristics of DDPM iterative diffusion; there is no need to calculate the model gradient during the verification stage through steps S3 and S4, and only black box permission requirements are required for the denoiser, which greatly reduces the white box permission requirements for model parameters and gradients required for the fingerprint scheme of the image restoration model in the existing technology; the fingerprint protection method for the image denoising diffusion probability model supports customized images in the verification process, for example, verification can be completed only by scanning the code, and the verification process is very intuitive and convenient; in terms of important indicators of the model fingerprint protection scheme, including uniqueness and robustness, quite good results have been achieved.

[0038] It is worth noting that in step S2, the protection target verification image is first Gradually add noise to obtain a set of verification images across the protection target Corresponding denoising diffusion probability model The noisy image of the denoiser performance boundary is used as the boundary crossing trajectory; and the image is verified based on the protection target. Corresponding denoising diffusion probability model The sampling process uses the model inversion to obtain the latent code component. In this embodiment, the latent code is Latent code, which is mapped to the protection target verification image. Model inversion is a method of optimizing model input by fixing model parameters and model output.

[0039] Preferably, in step S2, the protection target verification image As a denoising diffusion probability model Step 0 image , based on the denoising diffusion probability model The diffusion process of the image Gradually add a set of noise sequences , and obtain a set of noisy image sequences , where T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time;

[0040] By optimizing the noise sequence , so that the noisy image As t increases, it gradually crosses the protection target verification image Corresponding denoising diffusion probability model Denoiser The performance boundary area, and the trajectory image sequence at this time Verify image as the protection target Corresponding denoising diffusion probability model Denoiser boundary crossing trajectories.

[0041] In this embodiment, when calculating the image sequence During the process, the protection target verification image will be fixed and the protection target verification image Corresponding denoising diffusion probability model Denoiser , where the denoiser In fact, it is predicting the noise, which is the noise estimation model. The denoising model can usually be made by directly subtracting the estimated noise from the original image, so the noise estimation model is often regarded as a denoiser.

[0042] Optionally, in step S2, the protection target verification image Corresponding denoising diffusion probability model Denoiser , first give the noise at each diffusion time step t Initialize to get , get the initialized noise sequence ; for images Gradually add this set of noise sequences , get the initialized noisy image sequence ;

[0043] Formula (1): ;in, is a sample from a standard Gaussian distribution, is a noise image outside the Gaussian distribution. Distributions outside the Gaussian distribution include but are not limited to sampling of uniform distribution, Bernoulli distribution, etc. is the preset weight;

[0044] At each diffusion time step t, the protection target verification image and the denoising diffusion probability model corresponding to the protection target verification image are fixed Denoiser ,by As noise Initialization to optimize the noise , for the image Adding Noise After optimizing the noise, we get the image ;Finally, all the images corresponding to the diffusion time step t Composed set to obtain the verification image crossing the protection target Corresponding denoising diffusion probability model Denoiser Performance boundary region trajectory image sequence .

[0045] In this embodiment, the time step is timestep, and the denoising diffusion probability model DDPM often uses iterative denoising to achieve the generation process. It is common to give a total number of time steps. ,use Represents the time step of a specific step, that is .

[0046] Specifically, in step S2, the noise is optimized according to the gradient descent , λ represents the learning rate used in the optimization process; based on gradient descent optimization noise The loss function , formula (2): ;

[0047] in, are the preset weights, operators Refers to the calculation image The gradient can be achieved by simple pixel offset subtraction, T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time, t is the current diffusion time step, Denoiser Predicting noisy images The noise applied in to predict the optimized noise , Representation norm distance, Representation norm The distance, norm For all elements within the norm The sum of the absolute values, norm is the Euclidean distance, norm That is, each element in the norm The sum of the squares and then the square root.

[0048] By the formula and formula It can be seen that when t=1, the optimized noise It is easy to be denoised predicted, and when t=T, the denoiser Predicting optimized noise The process becomes very difficult, so the trajectory image sequence is converted into Defined as the protection target verification image Corresponding denoising diffusion probability model Denoiser boundary crossing trajectories.

[0049] This means that for each time step t, this solution needs to go through the optimization process, and the optimization can often be done by iterative gradient descent. , given the number of optimization iterations N and the learning rate used in the optimization process , for each time step t, first use the formula initialization , and then perform N gradient descents to reduce the noise Optimize.

[0050] For each time step t, the noise is optimized based on gradient descent The optimization process is isolated. Assuming that the number of optimization steps is N, for each t, we first use formula (1) to calculate And initialize it with , then cycle N times to optimize , the optimized loss function is formula (2), and each time the loss function is calculated, it can be updated by gradient descent ; Therefore, the total time complexity of the iteration is T*N times, and the noise calculated in t steps is The noise obtained with t+1 , theoretically they should be independent of each other.

[0051] It is worth noting that in step S2, the image By the current noise and images Denoised diffusion probability model The diffusion process yields: ,in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

[0052] Preferably, in step S2, the boundary crossing trajectory is used , through the denoising diffusion probability model The generation process of the noise perturbation added at each time step t is reversed , using pure noise and noise disturbance Composition Collection As a hidden code ;

[0053] Formula (3): ;in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in

[0054] The mean estimator is calculated as formula (4): ; , used for direct prediction ; , ,when is the sampling scheme of denoising diffusion probability model, when When is the sampling scheme of denoising diffusion probability model, this scheme mainly focuses on the sampling scheme of denoising diffusion probability model; is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in the scheduler and scheduler The parameters of the denoising diffusion probability model are predefined Scheduling strategy, scheduler and scheduler The parameters can be determined by customizing, using other existing scheduling schemes, or directly applying the default strategy of the denoising diffusion probability model. The scheduler here The setting needs to be consistent with the formula Scheduler in Same; the scheduler has T elements, each of which is a scalar (value) corresponding to each time step t, and t-1 is its subscript, so Represents the scheduler , the value used in the previous time step.

[0055] Denoising Diffusion Probabilistic Model The generation process has a total of T steps, usually T represents the largest subscript, and Indicates a clear image; Represents pure noise, denoising diffusion probability model The principle is to learn how to Generate clear images process. Therefore The randomness of will affect the certainty of the generated results, so it will be Perform inversion In addition, the generation sampling of the denoising diffusion probability model will introduce an additional random noise disturbance during the denoising process at each time step t. , The randomness introduced may also affect the certainty of the generated results. It is worth noting that In the original generation process, random Gaussian sampling is used, so it will be passed The inversion will Calculate and fix it. Therefore, in the denoising diffusion probability model In the process of generating , this scheme needs to determine the hidden code There are T+1 components, that is, .

[0056] Specifically, in step S3, the hidden code Input to the denoised diffusion probability model of suspected theft The generation process of , formula (5): , the image at t=1 As output image;

[0057] in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in ; when t=T, Hidden code in ;

[0058] Mean Estimator ; ; , ; is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

[0059] In this embodiment, the mean estimator in step S3 The scheduling strategy and the mean estimator in step S2 The scheduling strategy is the same; in step S3 The setting is in step S2 The settings are the same; in step S3 The setting is in step S2 The settings are the same; in step S3 The settings are the same as in step S2 In step S3, the denoising diffusion probability model of suspected theft is used. The samples are assigned to pure noise of Start iterative denoising and get a clear image after T steps , so the original sampling process only needs to focus on how to get That's it.

[0060] When using DDPM random sampling, the initial input and iterative injection are all random Gaussian noises. In the process of executing step S3, in order to ensure the denoising diffusion probability model of suspected theft To output the fingerprint verification image, you only need to execute the fingerprint code obtained in step S2 Replace its original and Specifically, the generation process of DDPM can be regarded as Step from pure noise Gradually remove noise to get a clear image process, and this process is from arrive arrive ...and finally process. The noisy image represents the intermediate process. When executing step S3 to perform fingerprint verification, the time step is from To begin with, there is fingerprint coding , so for formula (5), first calculate The scene, through and get , and with Gradually decrease to When , we can calculate , which is the final output image.

[0061] The generation process of DDPM is a gradual denoising process, so the DDPM model is essentially a denoiser, and the model parameters are also the parameters of the denoiser. Therefore, the denoising diffusion probability model for judging suspected theft is The authenticity of the theft is essentially to judge the denoising diffusion probability model of suspected theft Denoiser the authenticity of the theft.

[0062] It's worth noting that the fingerprint protection method for image denoising diffusion probability models is applicable not only to pixel-space denoising diffusion probability models (DDPMs), but also to latent-space diffusion models (LDMs). Experiments have shown that it is fully applicable to large-scale LDMs, including commonly used models such as Stable Diffusion and PixArt. The LDM (latent diffusion model) is a very common DDPM in AIGC, including models such as Stable Diffusion and PixArt. Its principle is to first compress and map the image to a low-dimensional latent space through an encoder. The DDPM is then learned in this latent space (diffusion of low-dimensional data samples and learning their distribution using a model). The decoder then maps the model back to the pixel space, saving resources. The DDPM model concept is applicable not only to images but also to other data domains such as video and audio. One problem with DDPM in the image domain is that the image space is large, making the mapping process complex. One approach is to first map the image to a low-dimensional latent space, which is smaller than the original pixel space and easier to fit and learn. The diffusion model, or DDPM, learned in this latent space is called the LDM. After learning the latent space distribution of an image, LDM then maps it back to pixel space through the decoder. LDM significantly reduces the complexity of fitting the original target domain distribution. Therefore, LDM can be simply viewed as first performing dimensionality reduction on the image, then using this reduced data to allow DDPM to learn its distribution.

[0063] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. A fingerprint protection method for image denoising diffusion probability model, characterized in that: The following steps are involved: S1: Obtaining protection target verification image ; The protected target verification image Built-in copyright information of the model owner; S2: Fixed protection target verification image and protection target verification image Corresponding denoising diffusion probability model , reverse protection target verification image Corresponding denoising diffusion probability model Hidden code Verify image as protection target Fingerprint code; S3: Obtaining a denoising diffusion probability model for suspected theft , the protection target verification image The fingerprint encoding is input into the denoising diffusion probability model of suspected theft The sampling process of the suspected theft is used to obtain the denoising diffusion probability model The corresponding output image; S4: Comparing the suspected theft denoising diffusion probability model The copyright information of the corresponding output image is verified with the protection target image Copyright information to determine the suspected theft of the denoising diffusion probability model The authenticity of the theft; wherein, when the denoising diffusion probability model of the suspected theft The copyright information of the corresponding output image is verified with the protection target image The denoising diffusion probability model indicates suspected theft when the copyright information is consistent The authenticity of the theft is true, otherwise it means the denoising diffusion probability model of the suspected theft The authenticity of the misappropriation is false.

2. The fingerprint protection method for image denoising diffusion probability model according to claim 1, characterized in that: In step S2, the protection target verification image is first Gradually add noise to obtain a set of verification images across the protection target Corresponding denoising diffusion probability model The noisy image of the denoiser performance boundary is used as the boundary crossing trajectory; and the image is verified based on the protection target. Corresponding denoising diffusion probability model The sampling process uses model inversion to obtain the latent coding component.

3. The fingerprint protection method for the image denoising diffusion probability model according to claim 2, characterized in that: In step S2, the protection target verification image As a denoising diffusion probability model Step 0 image , for the image Gradually add a set of noise sequences , and obtain a set of noisy image sequences , where T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time; By optimizing the noise sequence , so that the noisy image As t increases, it gradually crosses the protection target verification image Corresponding denoising diffusion probability model Denoiser The performance boundary area, and the trajectory image sequence at this time The protection target verification image Corresponding denoising diffusion probability model Denoiser boundary crossing trajectories.

4. The fingerprint protection method for image denoising diffusion probability model according to claim 3, characterized in that: In step S2, the protection target verification image Corresponding denoising diffusion probability model Denoiser , first give the noise at each diffusion time step t Initialize to get , get the initialized noise sequence ; ;in, is a sample from a standard Gaussian distribution, is a noise image outside the Gaussian distribution, is the preset weight; At each diffusion time step t, the protection target verification image and the denoising diffusion probability model corresponding to the protection target verification image are fixed Denoiser ,by As noise Initialization to optimize noise , for the image Adding Noise After optimizing the noise, we get the image ;Finally, all the images corresponding to the diffusion time step t Composed set to obtain the verification image crossing the protection target Corresponding denoising diffusion probability model Denoiser Performance boundary region trajectory image sequence .

5. The fingerprint protection method for image denoising diffusion probability model according to claim 4, characterized in that: In step S2, the noise is optimized according to gradient descent , λ represents the learning rate used in the optimization process; based on gradient descent optimization noise The loss function ; in, are the preset weights, operators Refers to the calculation image The gradient of T is the protection target verification image Corresponding denoising diffusion probability model The total number of steps of the preset diffusion time, t is the current diffusion time step, Denoiser Predicting noisy images The noise applied in to predict the optimized noise , Representation norm distance, Representation norm The distance, norm For all elements within the norm The sum of the absolute values, norm is the Euclidean distance.

6. The fingerprint protection method for image denoising diffusion probability model according to claim 5, characterized in that: In step S2, the image By the current noise and images Denoised diffusion probability model The diffusion process yields: ,in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

7. The fingerprint protection method for image denoising diffusion probability model according to claim 6, characterized in that: In step S2, the boundary crossing trajectory is used , through the denoising diffusion probability model The generation process of the noise perturbation added at each time step t is reversed , using pure noise and noise disturbance Composition Collection As a hidden code ; ;in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in The mean estimator is calculated as ; ; , , is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

8. The fingerprint protection method for image denoising diffusion probability model according to claim 7, characterized in that: In step S3, the hidden code Input to the denoised diffusion probability model of suspected theft The generation process , the image at t=1 As output image; in is the denoising diffusion probability model the mean estimator used in sampling, is the denoising diffusion probability model Variance schedule used in sampling An element in ; when t=T, Hidden code in ; Mean Estimator ; ; , ; is the denoising diffusion probability model Scheduler for iterative noise addition process An element in is the denoising diffusion probability model Scheduler for iterative noise addition process An element in .

Images