A zero watermark checking method and device for an image and a storage medium
By extracting latent space semantically robust features of images through a VAE and DDIM collaborative mechanism, and combining DWT and adversarial training, the low reliability of zero-watermark verification under latent space attacks and complex attacks in diffusion models is solved, and highly robust zero-watermark verification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING TECH UNIV
- Filing Date
- 2026-04-08
- Publication Date
- 2026-07-14
AI Technical Summary
Existing zero-watermark verification techniques have low verification reliability when facing latent space attacks and complex attacks in diffusion models. They also have weak semantic stability of pixel-domain features, lack deep semantic binding between features and watermarks, and lack consistency guarantees for cross-domain features.
A collaborative mechanism of variational autoencoder (VAE) and denoising diffusion implicit model (DDIM) is adopted to extract semantically robust features in the latent space. Frequency domain features are generated through discrete wavelet transform (DWT). A zero-watermark generation model with cross-domain consistency training and adversarial training is combined to achieve binding between cross-domain features and watermark semantics and improve robustness.
It significantly improves the robustness and verification reliability of zero watermark in the face of latent space attacks and complex attacks in diffusion models. By combining cross-domain consistency training and adversarial training, it achieves adaptive resistance to complex attacks and rapid verification.
Smart Images
Figure CN122390943A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a zero-watermark verification method, apparatus, and storage medium for images, belonging to the field of watermark verification technology. Background Technology
[0002] In early zero-watermark verification techniques, researchers mostly constructed copyright fingerprints based on pixel-domain features. For example, they extracted texture details, gradient distributions, or local statistical features of the image and fused them with watermark information using shallow methods such as XOR operations to achieve copyright authentication. In 2018, Jiren Zhu et al. proposed the HiDDeN model based on convolutional neural networks, which achieved breakthroughs in image quality and robustness against pixel-level attacks in traditional watermarking. However, when this approach, which relies on pixel-domain features, was directly applied to zero-watermarking technology, it exposed the limitation of extremely poor semantic stability of the shallow features extracted, making it unable to cope with the latent space deep reconstruction attacks of diffusion models. In 2019, Li Y et al. attempted to introduce shallow convolutional neural networks to extract pixel-domain deep features to optimize the resistance to traditional pixel-level attacks such as Gaussian noise and cropping. However, this method still did not escape the constraints of the pixel domain. When the diffusion model modified the latent space structure of the image, the pixel-domain features could not reflect the core semantics of the image, leading to the failure of copyright authentication. In addition, Xiyang Luo et al. used generative adversarial networks to optimize the fusion mechanism of features and watermarks in an attempt to solve the problem of loose coupling in traditional XOR operations. However, the inherent defect of unstable training of generative adversarial networks makes it difficult for features and watermarks to achieve deep semantic binding, and they are prone to decoupling under attacks. Even if subsequent studies introduce large kernel convolution to expand the receptive field of features, it is still limited to a single pixel domain and cannot take into account the consistency of latent space and frequency domain features. After the frequency domain energy distribution is smoothed by diffusion attacks, single domain features are very easy to become unstable.
[0003] In summary, existing zero-watermark verification still suffers from defects such as weak semantic stability of pixel domain features, lack of deep semantic binding between features and watermarks, and lack of consistency guarantee for cross-domain features, making it difficult to resist latent space attacks and complex attack combinations of diffusion models. This results in low verification reliability when facing latent space attacks and complex attacks of diffusion models. Summary of the Invention
[0004] The purpose of this invention is to provide a zero-watermark verification method, device, and storage medium for images, which solves the problem of low verification reliability in the face of diffusion model latent space attacks and complex attacks in the prior art.
[0005] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a zero-watermark verification method for images, comprising: The image to be verified is obtained, and the first latent space semantically robust features are extracted from the image through the VAE and DDIM collaborative mechanism. The first latent space semantically robust features are processed by DWT to generate the first frequency domain features, and the first latent space semantically robust features and the first frequency domain features are concatenated to form the first cross-domain feature group. The first cross-domain feature group is input into the trained zero-watermark generation model to obtain candidate zero-watermarks; A standard zero watermark is obtained, and the bit error rate of the candidate zero watermark and the standard zero watermark is calculated. If the bit error rate does not exceed a preset threshold, the image to be verified passes the zero watermark verification; otherwise, the image to be verified fails the zero watermark verification. The standard zero watermark is obtained by the following method: obtaining an initial zero watermark, obtaining a zero watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero watermark into the zero watermark optimization model to obtain the standard zero watermark.
[0006] Furthermore, the initial zero watermark is obtained through the following method: The original image and original watermark are obtained, and second latent space semantically robust features are extracted from the original image through a VAE and DDIM collaborative mechanism. The second latent space semantic robust features are subjected to DWT transformation to generate the second frequency domain features, and the second latent space semantic robust features and the second frequency domain features are concatenated to form the second cross-domain feature group. The original watermark is projected as an embedding vector, and the second cross-domain feature group and the embedding vector are input into the trained zero-watermark generation model to obtain the initial zero-watermark.
[0007] Based on the above-mentioned further technical solutions, the second latent space semantic robust features are extracted through the VAE and DDIM collaborative mechanism, the image is mapped to the latent space through VAE, and the semantics are enhanced through DDIM denoising, replacing the traditional pixel domain features, thereby improving the robustness of the initial zero watermark to the latent space attack of the diffusion model.
[0008] Furthermore, the initial zero-watermark is obtained by inputting the second cross-domain feature group and the embedded vector into the trained zero-watermark generation model, using the following formula: ; in, Indicates an initial zero watermark. Represents a symbolic function. This represents the trained ResNet network. This represents the second cross-domain feature group. This represents the embedding vector.
[0009] Based on the above-mentioned further technical solutions, the second cross-domain feature group and the embedding vector are input into the trained zero-watermark generation model. Semantic fusion is achieved through a pre-trained network, binding cross-domain features with watermark semantics, replacing the shallow fusion method, and avoiding the decoupling of features and watermark when facing attacks.
[0010] Furthermore, the step of extracting first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism includes: extracting first initial latent space features from the image to be verified through a trained VAE network, and optimizing the first initial latent space features through the DDIM method to obtain first latent space semantic robust features. The extraction of the first initial latent space features from the image to be verified using the trained VAE network is performed by the following formula: ; in, This represents the image to be verified. This represents the trained VAE network. Indicates the first initial latent space characteristics; The optimization of the first initial latent space features using the DDIM method includes: optimizing the first initial latent space features through deterministic iteration, completing the deterministic iteration after a preset number of steps, and obtaining the denoised first latent space semantically robust features. The iterative formula for the deterministic iterative denoising is: ; in, Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noise scheduling parameters for each iteration. Indicates the first The noise scheduling parameters for each iteration. Index representing the iteration step number, This represents the denoising network for DDIM; The training objective of the VAE network is to maximize the lower bound of evidence, and the expression for the lower bound of evidence is: ; in, Indicates the lower bound of evidence. This represents the latent variable obtained by encoding the image to be verified. Indicates finding hidden variables Follows distribution In the case of Expectations This represents the variational posterior distribution of the encoder output of the VAE network. This represents the likelihood distribution of the decoder in a VAE network. Denotes KL divergence, This represents the standard normal prior distribution. It represents the natural logarithm.
[0011] Based on the above-mentioned further technical solutions, the image is mapped to the latent space through VAE, and the semantics are enhanced by DDIM denoising, replacing the traditional pixel domain features, thereby improving the robustness of candidate zero watermarks to latent space attacks of diffusion models.
[0012] Furthermore, the step of performing DWT processing on the semantically robust features of the first latent space to generate the first frequency domain features includes: The semantically robust features of the first latent space are reshaped into a two-dimensional feature map. DWT processing is performed on each channel of the two-dimensional feature map to obtain low-frequency approximate sub-bands, horizontal detail sub-bands, vertical detail sub-bands, and diagonal detail sub-bands. All sub-bands are spliced and normalized to obtain the first frequency domain features.
[0013] Based on the aforementioned further technical solutions, DWT processing can provide a stable, interference-resistant, and highly discriminative feature foundation for subsequent zero-watermark construction. This transformation leverages the robustness of frequency domain features to conventional image processing while ensuring the transparency of the algorithm without modifying the original image. It also provides a good starting point for resisting geometric attacks and constructing hybrid domain features.
[0014] Furthermore, the expression for the loss function used during training of the zero-watermark optimization model is as follows: ； in, For the total loss, The first equilibrium hyperparameter, For the second equilibrium hyperparameter, For cross-domain consistency loss, To combat the losses; The expression for the cross-domain consistency loss is as follows: ; in, This represents the binary cross-entropy loss function. This represents the inverse process of the ResNet network. Indicates an initial zero watermark. This represents the second latent space semantically robust feature generated during the process of generating the initial zero watermark. This represents the second frequency domain feature generated during the process of generating the initial zero watermark; The expression for the adversarial loss is: ; in, This represents the second cross-domain feature group. This represents adversarial cross-domain feature sets used for adversarial training. Indicates the expectation. This represents the discriminator network in adversarial training. It represents the natural logarithm.
[0015] Based on the above-mentioned further technical solutions, by limiting the loss function, the combination of cross-domain consistency training and adversarial training is achieved as the same optimization step. Both the latent space-frequency domain feature semantic consistency is constrained by the shared encoder, and the diffusion attack is simulated by the adversarial network. The dual optimization yields the final robust zero watermark.
[0016] Furthermore, the bit error rate of the candidate zero watermark and the standard zero watermark is calculated using the following formula: ; in, Indicates bit error rate. Indicates the length of the zero watermark. This indicates the first standard zero watermark. 1 bit Indicates the first candidate for zero watermark 1 bit Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise.
[0017] Based on the aforementioned further technical solutions, zero-watermark verification is performed using the bit error rates of two zero-watermarks, effectively improving the accuracy, robustness, and reliability of zero-watermark verification. This mechanism leverages the redundancy and complementarity of the two watermarks to achieve adaptive resistance to complex attacks, providing richer decision-making basis for zero-watermark verification, while supporting the practical needs of blind extraction and rapid verification.
[0018] Secondly, the present invention provides a zero-watermark verification device for images, comprising: The semantic robust feature extraction module is configured to: acquire the image to be verified, and extract the first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism; The cross-domain feature concatenation module is configured to: perform DWT processing on the first latent space semantically robust features to generate the first frequency domain features, and concatenate the first latent space semantically robust features and the first frequency domain features to form the first cross-domain feature group; The zero-watermark extraction module is configured to: input the first cross-domain feature group into the trained zero-watermark generation model to obtain candidate zero-watermarks; The zero-watermark verification module is configured to: acquire a standard zero-watermark, calculate the bit error rate of the candidate zero-watermark and the standard zero-watermark, and if the bit error rate does not exceed a preset threshold, the image to be verified passes the zero-watermark verification; otherwise, the image to be verified fails the zero-watermark verification. The standard zero-watermark is obtained by the following method: acquiring an initial zero-watermark, acquiring a zero-watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero-watermark into the zero-watermark optimization model to obtain the standard zero-watermark.
[0019] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the zero-watermark verification method for images as described in any one of the first aspects.
[0020] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention provides a zero-watermark verification method, device, and storage medium for images. Through the collaboration of VAE and DDIM, VAE maps the image to the latent space, while DDIM denoises and enhances semantics, replacing traditional pixel-domain features and improving the robustness of the zero-watermark against latent space attacks of diffusion models. Semantic fusion is achieved through a trained zero-watermark generation model, binding cross-domain features and the semantic features of the zero-watermark, replacing shallow fusion methods and preventing features from decoupling from the zero-watermark when facing attacks. Cross-domain consistency training and adversarial training are combined, constraining semantic consistency between latent space and frequency domain features through a shared encoder, and simulating diffusion attacks through adversarial networks, thus doubly optimizing the robustness of the zero-watermark. The combination of these three techniques significantly improves the robustness of the zero-watermark against attacks, thereby solving the problem of low verification reliability in existing technologies when facing latent space attacks and complex attacks of diffusion models. Attached Figure Description
[0021] Figure 1 This is a flowchart of a zero-watermark verification method for images provided in an embodiment of the present invention; Figure 2 This is a flowchart of the standard zero-watermark generation process provided in the embodiments of the present invention. Detailed Implementation
[0022] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.
[0023] Example 1 like Figure 1 As shown, this embodiment provides a zero-watermark verification method for images, including: The image to be verified is obtained, and the first latent space semantic robust features are extracted from the image through the VAE and DDIM collaborative mechanism. VAE is Variational Autoencoder, and DDIM is Denoising Diffusion Implicit Models. The semantically robust features of the first latent space are processed by DWT to generate the first frequency domain features. The semantically robust features of the first latent space and the first frequency domain features are concatenated to form the first cross-domain feature group. Here, DWT stands for Discrete Wavelet Transform. The first cross-domain feature group is input into the trained zero-watermark generation model to obtain candidate zero-watermarks; A standard zero watermark is obtained, and the bit error rate of the candidate zero watermark and the standard zero watermark is calculated. If the bit error rate does not exceed a preset threshold, the image to be verified passes the zero watermark verification; otherwise, the image to be verified fails the zero watermark verification. The standard zero watermark is obtained by the following method: obtaining an initial zero watermark, obtaining a zero watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero watermark into the zero watermark optimization model to obtain the standard zero watermark.
[0024] This invention utilizes the collaboration of VAE and DDIM. VAE maps images to the latent space, while DDIM denoises and enhances semantics, replacing traditional pixel-domain features and improving the robustness of zero-watermark against latent space attacks of diffusion models. Semantic fusion is achieved through a trained zero-watermark generation model, binding cross-domain features and the semantic features of the zero-watermark, replacing shallow fusion methods and preventing features from decoupling from the zero-watermark when facing attacks. Cross-domain consistency training is combined with adversarial training, both constraining semantic consistency between latent space and frequency domain features through a shared encoder and simulating diffusion attacks through adversarial networks, thus doubly optimizing the robustness of the zero-watermark. The combination of these three techniques significantly improves the robustness of the zero-watermark against attacks, thereby solving the problem of low verification reliability in existing technologies when facing latent space attacks and complex attacks of diffusion models.
[0025] Example 2 This embodiment provides a zero-watermark verification method for images, specifically including the following steps S1 to S5.
[0026] Step S1: Obtain the image to be verified, and extract the first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism.
[0027] Step S1 specifically includes: Step S1.1: Extract the first initial latent space features from the image to be verified using the trained VAE network, that is, map the image to be verified to the latent space to obtain the first initial latent space features, the dimension of which is It can be done using the following formula: ; in, This represents the image to be verified. This represents the trained VAE network. This represents the first initial latent space feature.
[0028] The training objective of the VAE network is to maximize the lower bound of evidence, and the expression for the lower bound of evidence is: ; in, Indicates the lower bound of evidence. This represents the latent variable obtained by encoding the image to be verified. Indicates finding hidden variables Follows distribution In the case of Expectations This represents the variational posterior distribution of the encoder output of the VAE network. This represents the likelihood distribution of the decoder in a VAE network. Denotes KL divergence, This represents the standard normal prior distribution. It represents the natural logarithm.
[0029] Step S1.2: Optimize the first initial latent space features using the DDIM method, including: optimizing the first initial latent space features through deterministic iteration, completing the deterministic iteration after 200 steps, and obtaining the denoised first latent space semantically robust features.
[0030] The iterative formula for deterministic iterative denoising is as follows: ; in, Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noise scheduling parameters for each iteration. Indicates the first The noise scheduling parameters for each iteration. Index representing the iteration step number, This represents the denoising network for DDIM.
[0031] After 200 iterations, the semantically robust first latent space semantically robust features are obtained.
[0032] Step S2: Perform DWT processing on the first latent space semantic robust features to generate the first frequency domain features, and concatenate the first latent space semantic robust features and the first frequency domain features to form the first cross-domain feature group.
[0033] Specifically, the first frequency domain feature is obtained through the following method: the semantically robust features of the first latent space are reshaped into a two-dimensional feature map, and DWT processing is performed on each channel of the two-dimensional feature map to obtain low-frequency approximate sub-bands, horizontal detail sub-bands, vertical detail sub-bands and diagonal detail sub-bands. All sub-bands are spliced and normalized to obtain the first frequency domain feature.
[0034] Specifically, the first latent space semantically robust features and the first frequency domain features are concatenated into the first cross-domain feature group, which is done using the following formula: ; in, Indicates the first cross-domain feature group, This represents the semantic robustness of the first latent space. Represents the first frequency domain characteristic. This indicates a splicing operation.
[0035] Step S3: Combine the first cross-domain feature group The input is fed into the trained zero-watermark generation model to obtain candidate zero-watermarks.
[0036] In this embodiment, the zero-watermark generation model is specifically a RES fusion neural network.
[0037] Step S4: Obtain the initial zero watermark, and optimize it to obtain the standard zero watermark.
[0038] like Figure 2 As shown, step S4 can be further divided into steps S4.1 to S4.4.
[0039] Step S4.1: Obtain the original image and the original watermark, and extract the second latent space semantically robust features from the original image through the VAE and DDIM collaborative mechanism; Step S4.1 specifically includes: Step S4.1.1: Extract the second initial latent space features from the original image using the trained VAE network, according to the following formula: ; in, Represents the original image. This represents the trained VAE network. Indicates the second initial latent space characteristics; The VAE network used in step S4.1.1 is the same as the VAE network used in step S1.1.
[0040] Step S4.1.2: Optimize the second initial latent space features using the DDIM method, including: optimizing the second initial latent space features through deterministic iteration, completing the deterministic iteration after 200 steps, and obtaining the denoised second latent space semantically robust features.
[0041] The deterministic iteration method used in step S4.1.2 is the same as the deterministic iteration method used in step S1.2.
[0042] Step S4.2: Perform DWT transformation on the second latent space semantic robust features to generate the second frequency domain features, and concatenate the second latent space semantic robust features and the second frequency domain features to form the second cross-domain feature group.
[0043] Specifically, the second frequency domain features are obtained through the following method: the semantically robust features of the second latent space are reshaped into a new two-dimensional feature map, and DWT processing is performed on each channel of the new two-dimensional feature map to obtain the low-frequency approximate sub-band, horizontal detail sub-band, vertical detail sub-band and diagonal detail sub-band corresponding to the new two-dimensional feature map. All sub-bands are spliced and normalized to obtain the second frequency domain features.
[0044] Step S4.3: Project the original watermark into an embedding vector, and input the second cross-domain feature group and the embedding vector into the trained zero watermark generation model to obtain the initial zero watermark.
[0045] The original watermark is the original watermark after being encrypted using Logistic encryption.
[0046] The original binary watermark of length L is projected into an embedding vector through an embedding layer using the following formula: ; in, Indicates the original watermark. Indicates the embedding layer. This represents the embedding vector.
[0047] The initial zero-watermark is obtained by inputting the second cross-domain feature group and the embedded vector into the trained zero-watermark generation model, using the following formula: ; in, Indicates an initial zero watermark. This represents a sign function used to binarize a real-valued vector into {-1, 1} or {0, 1}. This refers to a trained ResNet network. In this embodiment, the zero-watermark generation model is specifically a ResNet network. This represents the second cross-domain feature group. This represents the embedding vector.
[0048] The trained zero-watermark generation model captures the deep semantic relationship between features and watermarks through a multi-head self-attention mechanism, and performs nonlinear fusion through a feedforward network to output the initial zero-watermark.
[0049] Step S4.4: Obtain the zero-watermark optimization model after cross-domain consistency training and adversarial training, and input the initial zero-watermark into the zero-watermark optimization model to obtain the standard zero-watermark. This step is to improve the robustness of the initial zero-watermark.
[0050] The expression for the loss function used during training of the zero-watermark optimization model, which has undergone cross-domain consistency training and adversarial training, is as follows: ; in, For the total loss, The first equilibrium hyperparameter, For the second equilibrium hyperparameter, For cross-domain consistency loss, To combat the losses.
[0051] The expression for cross-domain consistency loss is as follows: ; in, This represents the binary cross-entropy loss function. This represents the inverse process of the ResNet network, used to recover feature representations from a binarized watermark. Indicates an initial zero watermark. This represents the second latent space semantically robust feature generated during the process of generating the initial zero watermark. This represents the second frequency domain feature generated during the process of generating the initial zero watermark.
[0052] The expression for adversarial loss is: ; in, This represents the second cross-domain feature group. This represents adversarial cross-domain feature sets used for adversarial training. Indicates the expectation. This represents the discriminator network in adversarial training. It represents the natural logarithm.
[0053] Anti-training involves introducing an adversarial noise generator. and a discriminator Attacks such as generator-simulated diffusion models are used to target... Applying a perturbation yields The discriminator attempts to distinguish between the original feature set and the attacked feature set. Through adversarial training, the binary zero-watermark generation process becomes invariant to such perturbations.
[0054] After obtaining the standard zero watermark, the standard zero watermark, the image's unique identifier, and copyright information can be stored in a third-party copyright library and retrieved directly when needed.
[0055] After step S4.4, the consistency of watermark decoding of latent space-frequency domain features is constrained through cross-domain consistency training, the watermark's resistance to attacks such as diffusion models is improved through adversarial training, and the network parameters are jointly iteratively updated through the total loss function, finally obtaining a standard zero watermark with robustness.
[0056] Step S5: Calculate the bit error rate of the candidate zero watermark and the standard zero watermark. If the bit error rate does not exceed the preset threshold, the image to be verified passes the zero watermark verification; otherwise, the image to be verified fails the zero watermark verification.
[0057] The bit error rate of the candidate zero watermark and the standard zero watermark is calculated using the following formula: ; in, Indicates bit error rate. Indicates the length of the zero watermark. This indicates the first standard zero watermark. 1 bit Indicates the first candidate for zero watermark 1 bit Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise.
[0058] In this embodiment, the preset threshold for the bit error rate is 5%.
[0059] If the image to be verified passes the zero-watermark verification, then the copyright ownership of the image to be verified is valid; otherwise, the copyright ownership of the image to be verified is invalid.
[0060] Example 3 This embodiment provides a zero-watermark verification device for images, including: The semantic robust feature extraction module is configured to: acquire the image to be verified, and extract the first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism; The cross-domain feature concatenation module is configured to: perform DWT processing on the first latent space semantically robust features to generate the first frequency domain features, and concatenate the first latent space semantically robust features and the first frequency domain features to form the first cross-domain feature group; The zero-watermark extraction module is configured to: input the first cross-domain feature group into the trained zero-watermark generation model to obtain candidate zero-watermarks; The zero-watermark verification module is configured to: acquire a standard zero-watermark, calculate the bit error rate of the candidate zero-watermark and the standard zero-watermark, and if the bit error rate does not exceed a preset threshold, the image to be verified passes the zero-watermark verification; otherwise, the image to be verified fails the zero-watermark verification. The standard zero-watermark is obtained by the following method: acquiring an initial zero-watermark, acquiring a zero-watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero-watermark into the zero-watermark optimization model to obtain the standard zero-watermark.
[0061] Example 4
[0062] This embodiment provides a computer-readable storage medium storing a computer program / instructions thereon. When the computer program / instructions are executed by a processor, they implement the steps of the zero-watermark verification method for images provided in Embodiment 1: The image to be verified is obtained, and the first latent space semantically robust features are extracted from the image through the VAE and DDIM collaborative mechanism. The first latent space semantically robust features are processed by DWT to generate the first frequency domain features, and the first latent space semantically robust features and the first frequency domain features are concatenated to form the first cross-domain feature group. The first cross-domain feature group is input into the trained zero-watermark generation model to obtain candidate zero-watermarks; A standard zero watermark is obtained, and the bit error rate of the candidate zero watermark and the standard zero watermark is calculated. If the bit error rate does not exceed a preset threshold, the image to be verified passes the zero watermark verification; otherwise, the image to be verified fails the zero watermark verification. The standard zero watermark is obtained by the following method: obtaining an initial zero watermark, obtaining a zero watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero watermark into the zero watermark optimization model to obtain the standard zero watermark.
[0063] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0064] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0067] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A zero-watermark verification method for images, characterized in that, include: The image to be verified is obtained, and the first latent space semantically robust features are extracted from the image through the VAE and DDIM collaborative mechanism. The first latent space semantically robust features are processed by DWT to generate the first frequency domain features, and the first latent space semantically robust features and the first frequency domain features are concatenated to form the first cross-domain feature group. The first cross-domain feature group is input into the trained zero-watermark generation model to obtain candidate zero-watermarks; A standard zero watermark is obtained, and the bit error rate of the candidate zero watermark and the standard zero watermark is calculated. If the bit error rate does not exceed a preset threshold, the image to be verified passes the zero watermark verification; otherwise, the image to be verified fails the zero watermark verification. The standard zero watermark is obtained by the following method: obtaining an initial zero watermark, obtaining a zero watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero watermark into the zero watermark optimization model to obtain the standard zero watermark.
2. The zero-watermark verification method for images according to claim 1, characterized in that, The initial zero watermark is obtained through the following method: The original image and original watermark are obtained, and second latent space semantically robust features are extracted from the original image through a VAE and DDIM collaborative mechanism. The second latent space semantic robust features are subjected to DWT transformation to generate the second frequency domain features, and the second latent space semantic robust features and the second frequency domain features are concatenated to form the second cross-domain feature group. The original watermark is projected as an embedding vector, and the second cross-domain feature group and the embedding vector are input into the trained zero-watermark generation model to obtain the initial zero-watermark.
3. The zero-watermark verification method for images according to claim 2, characterized in that, The initial zero-watermark is obtained by inputting the second cross-domain feature group and the embedded vector into the trained zero-watermark generation model, using the following formula: ； in, Indicates an initial zero watermark. Represents a symbolic function. This represents the trained ResNet network. This represents the second cross-domain feature group. This represents the embedding vector.
4. The zero-watermark verification method for images according to claim 1, characterized in that, The step of extracting first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism includes: extracting first initial latent space features from the image to be verified through a trained VAE network, and optimizing the first initial latent space features through the DDIM method to obtain first latent space semantic robust features. The extraction of the first initial latent space features from the image to be verified using the trained VAE network is performed by the following formula: ； in, This represents the image to be verified. This represents the trained VAE network. Indicates the first initial latent space characteristics; The optimization of the first initial latent space features using the DDIM method includes: optimizing the first initial latent space features through deterministic iteration, completing the deterministic iteration after a preset number of steps, and obtaining the denoised first latent space semantically robust features. The iterative formula for the deterministic iterative denoising is: ； in, Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noisy characteristics of step-by-step iteration Indicates the first The noise scheduling parameters for each iteration. Indicates the first The noise scheduling parameters for each iteration. Index representing the iteration step number, This represents the denoising network for DDIM; The training objective of the VAE network is to maximize the lower bound of evidence, and the expression for the lower bound of evidence is: ； in, Indicates the lower bound of evidence. This represents the latent variable obtained by encoding the image to be verified. Indicates finding hidden variables Follows distribution In the case of Expectations This represents the variational posterior distribution of the encoder output of the VAE network. This represents the likelihood distribution of the decoder in a VAE network. Denotes KL divergence, This represents the standard normal prior distribution. It represents the natural logarithm.
5. The zero-watermark verification method for images according to claim 1, characterized in that, The step of performing DWT processing on the semantically robust features of the first latent space to generate the first frequency domain features includes: The semantically robust features of the first latent space are reshaped into a two-dimensional feature map. DWT processing is performed on each channel of the two-dimensional feature map to obtain low-frequency approximate sub-bands, horizontal detail sub-bands, vertical detail sub-bands, and diagonal detail sub-bands. All sub-bands are spliced and normalized to obtain the first frequency domain features.
6. The zero-watermark verification method for images according to claim 1, characterized in that, The expression for the loss function used during training of the zero-watermark optimization model is as follows: ； in, For the total loss, The first equilibrium hyperparameter, For the second equilibrium hyperparameter, For cross-domain consistency loss, To combat the losses; The expression for the cross-domain consistency loss is as follows: ； in, This represents the binary cross-entropy loss function. This represents the inverse process of the ResNet network. Indicates an initial zero watermark. This represents the second latent space semantically robust feature generated during the process of generating the initial zero watermark. This represents the second frequency domain feature generated during the process of generating the initial zero watermark; The expression for the adversarial loss is: ； in, This represents the second cross-domain feature group. This represents adversarial cross-domain feature sets used for adversarial training. Indicates the expectation. This represents the discriminator network in adversarial training. It represents the natural logarithm.
7. The zero-watermark verification method for images according to claim 1, characterized in that, The bit error rate of the candidate zero watermark and the standard zero watermark is calculated using the following formula: ； in, Indicates bit error rate, Indicates the length of the zero watermark. This indicates the first standard zero watermark. 1 bit Indicates the first candidate for zero watermark 1 bit Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise.
8. A zero-watermark verification device for images, characterized in that, include: The semantic robust feature extraction module is configured to: acquire the image to be verified, and extract the first latent space semantic robust features from the image to be verified through the VAE and DDIM collaborative mechanism; The cross-domain feature concatenation module is configured to: perform DWT processing on the first latent space semantically robust features to generate the first frequency domain features, and concatenate the first latent space semantically robust features and the first frequency domain features to form the first cross-domain feature group; The zero-watermark extraction module is configured to: input the first cross-domain feature group into the trained zero-watermark generation model to obtain candidate zero-watermarks; The zero-watermark verification module is configured to: acquire a standard zero-watermark, calculate the bit error rate of the candidate zero-watermark and the standard zero-watermark, and if the bit error rate does not exceed a preset threshold, the image to be verified passes the zero-watermark verification; otherwise, the image to be verified fails the zero-watermark verification. The standard zero-watermark is obtained by the following method: acquiring an initial zero-watermark, acquiring a zero-watermark optimization model that has undergone cross-domain consistency training and adversarial training, and inputting the initial zero-watermark into the zero-watermark optimization model to obtain the standard zero-watermark.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the zero-watermark verification method for images as described in any one of claims 1 to 7.