Trajectory perturbation image watermarking processing method based on diffusion model and application thereof

Abstract

The application relates to the technical field of intelligent image processing, in particular to a trajectory perturbation image watermark processing method based on a diffusion model and application thereof. Deterministic evolution trajectories are determined, detection uncertainty caused by random sampling of a traditional diffusion model is eliminated, watermark detection can be completed through single backtracking, and thus detection efficiency is improved; watermark information is introduced into an evolution process as a dynamic constraint, endogenous coupling between the watermark and image content is realized through perturbation instantaneous velocity vectors, the watermark signal and image semantics are deeply fused, and strong robustness is achieved against malicious attacks such as cutting, compression and noise interference; in the detection stage, an instantaneous velocity bias of a trajectory is generated through prediction, a copyright fingerprint feature vector is constructed, and convenient detection of the watermark is realized through similarity discrimination between the copyright fingerprint feature vector and a watermark information vector. The application aims to solve the problems of how to overcome watermark detection instability caused by randomness of generated trajectories and low coupling degree between the watermark and image content.

Description

Technical Field

[0001] This application relates to the field of intelligent image processing technology, and in particular to a trajectory perturbation image watermarking method based on a diffusion model and its application. Background Technology

[0002] With the rapid development of generative artificial intelligence technology, diffusion models, with their superior modeling capabilities for complex data distributions, have become the mainstream technology in the field of image generation. In the field of digital copyright protection, the use of diffusion models to embed digital watermarks has become a research hotspot.

[0003] Existing solutions for embedding digital watermarks using diffusion models still have significant limitations in practical applications. On the one hand, most solutions employ secondary processing using an external encoding module after image generation, resulting in the watermark signal remaining only on the pixel surface with low coupling to image features, making it susceptible to failure under fine-grained cropping, color adjustment, or resampling attacks. On the other hand, the sampling logic of mainstream diffusion models essentially relies on the gradual injection of random noise. This non-deterministic generation path makes the watermark signal susceptible to random interference when reverse-engineered at the detection end, making it difficult to achieve stable and reliable watermark detection and tracing.

[0004] In view of this, this application proposes a new image watermarking processing method, which aims to solve the problem of random interference in watermark detection and achieve highly robust endogenous watermark generation and deterministic detection. Summary of the Invention

[0005] The main objective of this application is to provide a trajectory perturbation image watermarking method based on a diffusion model, aiming to overcome the problems of unstable watermark detection caused by the randomness of generated trajectories and low coupling between watermark and image content.

[0006] To achieve the above objectives, this application provides a trajectory perturbation image watermarking method based on a diffusion model, the method comprising: In the watermark generation stage, the watermark information vector is introduced into the construction process of the deterministic evolution trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information. The deterministic evolution trajectory represents the process of the image to be embedded with watermark evolving towards a noise distribution. In the watermark detection stage, Gaussian noise seeds are obtained in the image to be detected. A deterministic observation path corresponding to the image to be detected is determined based on the Gaussian noise seeds. An instantaneous velocity bias is predicted along the deterministic observation path and a copyright fingerprint feature vector is generated. Watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seeds.

[0007] Optionally, the watermark information vector is introduced into the construction process of the deterministic evolution trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information, specifically including: S10, map the preset watermark sequence into an original feature vector that satisfies the preset spatial dimension of the diffusion model, and modulate the original feature vector into a conditional feature vector that satisfies the image evolution trajectory condition through a nonlinear activation function. Furthermore, the image to be embedded with the watermark is mapped to the latent space to generate initial image features, and the initial image features are linearly interpolated with standard Gaussian noise according to a preset sampling time distribution to construct a deterministic evolution trajectory from the target image distribution to the initial noise distribution; S20, based on the conditional feature vector, the instantaneous velocity vector of the deterministic evolution trajectory is perturbed, and the initial image features are evolved along the deterministic evolution trajectory that has been deflected after the perturbation, so that the watermark sequence can achieve synchronous endogenous coupling during the image generation process, and image features carrying endogenous watermarks are obtained. S30, the image features carrying the endogenous watermark are mapped to the pixel domain to generate the endogenous watermark image.

[0008] Optionally, in step S20, the initial image features are evolved along the deterministic evolution trajectory after perturbation by perturbing the instantaneous velocity vector in the deterministic evolution trajectory based on the conditional feature vector. This specifically includes: S21, use the sine-cosine coding function to process the current sampling time step. Perform high-dimensional mapping to obtain time-step embedding vectors Embed the time step into the vector With the conditional feature vector Feature fusion is performed, and an adaptive transformation matrix is ​​used. With the corresponding bias term Dynamic response parameters for fitting the denoised trajectory:

[0009] In the formula, and These are the scaling factor and offset factor used to adjust the evolution trajectory, respectively; S22, the dynamic response parameters are applied between the self-attention operator and the feedforward network of the Transformer block to form a controlled vector field function dynamically guided by the watermark. ; S23, the controlled vector field function is subjected to inverse integral evolution from the time step direction through an ordinary differential equation solver; S24, During the integration process, based on the current time step With conditional eigenvectors By fitting the bias of the instantaneous velocity vector in real time, the latent features are driven to evolve along a controlled trajectory, thereby achieving endogenous coupling of the watermark signal during image generation.

[0010] Optionally, step S23 includes: The feature evolution at each sampling step follows the following linear integral operator:

[0011] In the formula, For single-step evolution increments, Used to characterize the current feature Conditional eigenvectors and time step Evolution rate under common constraints.

[0012] Optionally, predicting the instantaneous velocity bias along the deterministic observation path and generating a copyright fingerprint feature vector specifically includes: S40, invoke the pre-trained encoder to map the image to be detected to the latent space and obtain its corresponding suspected latent feature representation. ; S50, invoke the pre-stored initial Gaussian noise seed, and establish the representation of the suspected latent features according to the linear interpolation formula. Pointing to the latent feature representation as the noise endpoint Deterministic observation path :

[0013] In the formula, Total evolution time For time steps; S60 will define the characteristics of sampling points along the deterministic observation path. The temporal embedding vector corresponding to the features of the sampling points and conditional eigenvectors The input is fed into a controlled denoising network with frozen weights. The instantaneous velocity vector output by the controlled denoising network is compared with the linear velocity vector of the deterministic observation path itself to extract the residual watermark dynamic bias component. S70 performs path integral accumulation on the bias components extracted from each sampling point in the time series, and then performs global pooling to construct a copyright fingerprint feature vector that is consistent with the preset watermark control feature dimension.

[0014] Optionally, the expression for the copyright fingerprint feature vector is:

[0015] In the formula, This represents the total number of discrete sampling steps; Represents the instantaneous velocity vector; For single-step evolution increments; Represents latent features; Indicates potential characteristics of a suspect; This indicates the total evolution time.

[0016] Optionally, watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seed, specifically including: Calculate the cosine similarity between the fingerprint feature vector and the preset watermark control feature; If the cosine similarity is greater than a preset threshold, then the image to be detected is determined to contain endogenous watermark trajectory features.

[0017] In addition, to achieve the above objectives, this application also provides an application of the trajectory perturbation image watermarking processing method based on the diffusion model as described in any of the preceding claims, which is applied to image watermark generation and / or image watermark detection.

[0018] In addition, to achieve the above objectives, this application also provides a computer system, the computer system comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps of the trajectory perturbation image watermarking method based on the diffusion model as described in any of the preceding claims.

[0019] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the trajectory perturbation image watermarking method based on the diffusion model as described in any of the preceding claims.

[0020] This application has at least the following beneficial effects: 1. By using deterministic evolutionary trajectories, the detection uncertainty caused by random sampling in traditional diffusion models is eliminated, allowing watermark detection to be completed with only a single backtracking, thereby improving detection efficiency; 2. Watermark information is introduced into the evolution process as a dynamic constraint. Endogenous coupling between watermark and image content is achieved through perturbation of instantaneous velocity vector, enabling deep fusion of watermark signal and image semantics. This provides strong robustness against malicious attacks such as cropping, compression, and noise interference. 3. In the detection phase, a copyright fingerprint feature vector is constructed by predicting the instantaneous velocity bias of the generated trajectory. The watermark is then conveniently detected by judging the similarity between the copyright fingerprint feature vector and the watermark information vector. Attached Figure Description

[0021] Figure 1This is a schematic flowchart of the trajectory perturbation image watermarking processing method based on the diffusion model involved in the embodiments of this application; Figure 2 This is a schematic diagram of the overall network architecture of the trajectory perturbation image based on the diffusion model involved in the embodiments of this application; Figure 3 This is a schematic diagram of the architecture of the controlled denoising network involved in the embodiments of this application; Figure 4 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.

[0022] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0024] First Embodiment Reference Figure 1 and Figure 2 This embodiment provides a trajectory perturbation image watermarking method based on a diffusion model. The processing method involved in this embodiment includes two stages: a watermark generation stage and a watermark detection stage. The method includes the following steps: S100, in the watermark generation stage, the watermark information vector is introduced into the construction process of the deterministic evolution trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information. The deterministic evolution trajectory represents the process of the watermark image to be embedded evolving towards a noise distribution. In this embodiment, a deterministic evolutionary trajectory refers to a unique and reproducible path in the latent space of image generation that evolves from an initial noise state to a final image feature state. Unlike traditional diffusion models that rely on random sampling paths, deterministic evolutionary trajectories eliminate random noise interference during the generation process, ensuring that each initial noise state uniquely corresponds to a final image feature state.

[0025] In the dynamics of image generation, watermark information is used to perturb the instantaneous velocity vector of the evolution trajectory. This perturbation causes a small but controllable bias in the originally standard evolution trajectory. Because this bias occurs synchronously during the gradual generation of image features, the watermark information is deeply embedded in the semantic features of the image, rather than merely remaining on the pixel surface. This fusion method makes the watermark and image content an organic whole, i.e., an endogenous watermarked image. The endogenous characteristics of the endogenous watermarked image give the watermark extremely high robustness. Even if the image is subjected to attacks such as cropping, compression, or color transformation, as long as the core semantic features of the image are preserved, the watermark information is difficult to remove.

[0026] In some alternative implementations, the watermark information vector is a 64-bit binary bitstream of copyright information.

[0027] Specifically, it includes: S10, map the preset watermark sequence into an original feature vector that satisfies the preset spatial dimension of the diffusion model, and modulate the original feature vector into a conditional feature vector that satisfies the image evolution trajectory condition through a nonlinear activation function. For example, the preset copyright identity information ,in This identifies the original dimension of the information. A learnable linear projection matrix is ​​used. With bias term Perform a spatial transformation to map the discrete identity information to the latent feature space of the diffusion model, obtaining the original feature vector. :

[0028] In the formula, , For potential spatial feature dimensions.

[0029] The original feature vector The input nonlinear activation layer is used for feature reshaping to determine the final conditional feature vector used to regulate the evolutionary trajectory. :

[0030] In the formula, It is a nonlinear function, specifically represented as:

[0031] Furthermore, the image to be embedded with the watermark is mapped to the latent space to generate initial image features, and the initial image features are linearly interpolated with standard Gaussian noise according to a preset sampling time distribution to construct a deterministic evolution trajectory from the target image distribution to the initial noise distribution; For example, the encoder operator in a pre-trained variational autoencoder is used to map the original image to the latent space to obtain an initial latent feature representation. Initial latent feature representation is obtained by sampling according to the standard normal distribution. Dimensionally consistent initial Gaussian noise .

[0032] See also Figure 3 The schematic diagram of the controlled denoising network shown illustrates the initial Gaussian noise. As the maximum time step Potential feature representation at the location Based on the linear flow matching framework, we construct a representation from the initial latent features. Towards initial Gaussian noise A deterministic forward path of evolution. For any time step... The current features obtained by defining its linear standard deterministic evolution trajectory. for:

[0033] when from continuously increasing to hour, From containing complete semantic information Linear offset, ultimately in When the unbiased latent feature representation is fully reached... .

[0034] S20, based on the conditional feature vector, the instantaneous velocity vector of the deterministic evolution trajectory is perturbed, and the initial image features are evolved along the deterministic evolution trajectory that has been deflected after the perturbation, so that the watermark sequence can achieve synchronous endogenous coupling during the image generation process, and image features carrying endogenous watermarks are obtained. Specifically, it includes: S21, use the sine-cosine coding function to process the current sampling time step. Perform high-dimensional mapping to obtain time-step embedding vectors Embed the time step into the vector With the conditional feature vector Feature fusion is performed, and an adaptive transformation matrix is ​​used. With the corresponding bias term Dynamic response parameters for fitting the denoised trajectory:

[0035] In the formula, and These are the scaling factor and offset factor used to adjust the evolution trajectory, respectively; S22, the dynamic response parameters are applied between the self-attention operator and the feedforward network of the Transformer block to form a controlled vector field function dynamically guided by the watermark. ; S23, the controlled vector field function is subjected to inverse integral evolution from the time step direction through an ordinary differential equation solver; S24, During the integration process, based on the current time step With conditional eigenvectors By fitting the bias of the instantaneous velocity vector in real time, the latent features are driven to evolve along a controlled trajectory, thereby achieving endogenous coupling of the watermark signal during image generation.

[0036] S30, the image features carrying the endogenous watermark are mapped to the pixel domain to generate the endogenous watermark image.

[0037] For example, when the inverse integral iterates The next execution until the termination time step At that time, latent feature representations containing watermark perturbation features are obtained. .

[0038] Decoder operators in variational autoencoders The latent features are mapped back to pixel space to output the final watermarked image. :

[0039] Furthermore, and optionally, in this embodiment, we also construct a joint loss function for an ordinary differential equation solver based on conditional flow matching. The calculation formula is as follows:

[0040] In the formula, The mathematical expectation operator refers to the average loss calculated over the entire training dataset and time step distribution. It is the square of the L2 norm.

[0041] S200, in the watermark detection stage, a Gaussian noise seed in the image to be detected is obtained, a deterministic observation path corresponding to the image to be detected is determined according to the Gaussian noise seed, an instantaneous velocity bias is predicted along the deterministic observation path and a copyright fingerprint feature vector is generated, and watermark detection is performed according to the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seed.

[0042] In this embodiment, the detection phase is the reverse process of the generation phase. Since the generation phase constructs a deterministic evolutionary trajectory, the detection phase can precisely backtrack along this deterministic path. During this process, this embodiment compares the actual evolutionary state with the baseline evolutionary state (i.e., the standard state without watermark constraints). Because dynamic constraints are introduced in the generation phase, the actual evolutionary trajectory develops a controllable bias, which remains as a dynamic deviation during the backtracking process. By accumulating these deviations, a fingerprint vector reflecting the watermark features can be constructed. This detection method based on trajectory deviation accumulation avoids the detection instability problem caused by the irreversibility of random sampling paths in traditional diffusion model watermark detection, enabling watermark feature extraction in a single backtracking.

[0043] Specifically and optionally, predicting the instantaneous velocity bias and generating a copyright fingerprint feature vector along the deterministic observation path specifically includes: S40, invoke the pre-trained encoder to map the image to be detected to the latent space and obtain its corresponding suspected latent feature representation. ; In some alternative implementations, the pre-trained encoder is selected as a variational autoencoder.

[0044] S50, invoke the pre-stored initial Gaussian noise seed, and establish the representation of the suspected latent features according to the linear interpolation formula. Pointing to the latent feature representation as the noise endpoint Deterministic observation path :

[0045] In the formula, Total evolution time For time steps; S60 will define the characteristics of sampling points along the deterministic observation path. The temporal embedding vector corresponding to the features of the sampling points and conditional eigenvectors The input is fed into a controlled denoising network with frozen weights. The instantaneous velocity vector output by the controlled denoising network is compared with the linear velocity vector of the deterministic observation path itself to extract the residual watermark dynamic bias component. S70 performs path integral accumulation on the bias components extracted from each sampling point in the time series, and then performs global pooling to construct a copyright fingerprint feature vector that is consistent with the preset watermark control feature dimension.

[0046] Further, and optionally, the expression for the copyright fingerprint feature vector is:

[0047] In the formula, This represents the total number of discrete sampling steps; Represents the instantaneous velocity vector; For single-step evolution increments; Represents latent features; Indicates potential characteristics of a suspect; This indicates the total evolution time.

[0048] Specifically and optionally, watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seed, specifically including: Calculate the cosine similarity between the fingerprint feature vector and the preset watermark control feature; If the cosine similarity is greater than a preset threshold, then the image to be detected is determined to contain endogenous watermark trajectory features.

[0049] It should be noted that the endogenous watermark trajectory feature refers to the endogenous watermark obtained based on the watermark generation stage in this embodiment.

[0050] In the technical solution provided in this embodiment, the deterministic evolution trajectory eliminates the detection uncertainty caused by random sampling in traditional diffusion models, enabling watermark detection to be completed with only a single backtracking, thereby improving detection efficiency. On the other hand, watermark information is introduced into the evolution process as a dynamic constraint, and the watermark and image content are intrinsically coupled through perturbation of instantaneous velocity vectors, enabling deep fusion of watermark signals and image semantics, and exhibiting strong robustness against malicious attacks such as cropping, compression, and noise interference. Furthermore, in the detection stage, a copyright fingerprint feature vector is constructed by predicting the instantaneous velocity bias of the generated trajectory, and convenient watermark detection is achieved by judging the similarity between the copyright fingerprint feature vector and the watermark information vector.

[0051] Furthermore, as an implementation scheme, this embodiment provides an application of the trajectory perturbation image watermarking processing method based on the diffusion model as described in any of the above claims, specifically applied to image watermark generation and / or image watermark detection.

[0052] Second Embodiment Based on the trajectory perturbation image watermarking method described in the first embodiment, this embodiment focuses on achieving end-to-end endogenous watermark coupling during the image generation process using a controlled vector field in the watermark generation stage. Specifically: Convert the preset watermark sequence into a high-dimensional control feature vector; In this embodiment, the watermark information is a 64-bit binary bitstream of copyright information.

[0053] Using a pre-defined linear projection layer, the input 64-dimensional watermark sequence is... Directly mapped to high-dimensional conditional feature vectors that match the dimensions of the network's latent feature space. :

[0054] In the formula, The dimension is The weight matrix, This corresponds to the 1024-dimensional bias term.

[0055] The identity feature vector The input nonlinear activation layer is used for feature reshaping to determine the final 1024-dimensional conditional feature vector used to regulate the evolutionary trajectory. :

[0056] In the formula, It is a nonlinear function, specifically represented as:

[0057] Obtain the initial noise state of the original image. In this embodiment, the original image... for The original image has RGB three channels. The encoder part of a pre-trained variational autoencoder is used to process the original image. Mapping to a latent feature space with 4 feature channels yields the initial latent features of the original image. Its dimensions are Based on sampling according to the standard normal distribution, initial latent features are obtained. Dimensionally consistent initial Gaussian noise Set the total evolution duration. The initial Gaussian noise As the maximum time step Potential feature representation at the location In this embodiment, the evolutionary trajectory is defined over a continuous interval. Above, among which, Corresponding to the initial Gaussian noise distribution, Distribution of potential watermark features corresponding to the target.

[0058] Set the total number of iterations to .

[0059] Calculate the single-step evolution increment based on the linear discretization principle. Constructing discrete sequences ,in, , This is used to guide the network to perform corresponding tangent vector predictions at different generation stages. For the current time step... Perform preprocessing to transform it into a temporal embedding vector. This allows scalar time information to be integrated with a 1024-dimensional conditional eigenvector. Nonlinear coupling is performed in the feature space.

[0060] Representing latent features Conditional eigenvectors and temporal embedding vector Input a pre-trained controlled denoising network, the network being based on a diffusion transformer architecture, with the specific structure as follows: Figure 2 As shown: The controlled denoising network described in this embodiment contains 32 cascaded and stacked Transformer Blocks. Structurally, the network is divided into a backbone feature flow and a lateral control flow. In the backbone feature flow, latent feature representations... Store the current features in the graph The latent feature register unit. In each iteration step, The feature vectors are sequentially processed through each TransformerBlock. The output of the previous Block serves as the input to the next Block, achieving layer-by-layer refinement of the features. In the lateral control flow, the conditional feature vectors are... and temporal embedding vector Parallel input is fed into the adaptive layer normalization unit within each block level.

[0061] Through adaptive transformation matrix With the corresponding bias term Dynamically fit a set of scaling factors for the four channels of the latent features. and offset coefficient :

[0062] The adaptive layer normalization unit generates an intermediate feature stream modulated by the watermark signal by performing an affine transformation on the input latent features. :

[0063] In the formula, For the current features, Presentation layer normalization operation.

[0064] intermediate feature stream The input multi-head self-attention module performs spatial modeling and passes it to the next-level transformer block. After the feature flow passes through the last-level transformer block, the network predicts and outputs the instantaneous velocity vector dynamically guided by the watermark at the current moment. . Real-time characterization of current features Watermarked vector and time step Evolution rate under common constraints.

[0065] Furthermore, Input this into the ODE solver, and combine it with the preset step size. Solving for the latent feature states at the next time step :

[0066] And update the latent feature state It is fed back to the first transformer block as input for the next iteration.

[0067] When the inverse integral iterates Next, execute until the termination time step At that time, latent feature representations containing watermark perturbation features are obtained. .

[0068] Call the decoder unit in the pre-trained variational autoencoder , with dimensions as Watermarking potential features Restored to dimension Watermarked images :

[0069] Third Embodiment Based on the trajectory perturbation image watermarking method described in the first embodiment, the focus of this embodiment is on determining whether the watermark correlation exceeds a threshold during the detection stage, thereby determining the copyright information.

[0070] Obtain the latent feature state of the suspect image to be detected.

[0071] In this embodiment, the specifications of the suspect image are consistent with those of the first embodiment. RGB three-channel image .

[0072] The encoder operator in the same pre-trained variational autoencoder as in the first embodiment is invoked to process the suspect image. Mapping to the latent space, we get a dimension of Suspect potential characteristics .

[0073] Set the total evolution duration line .

[0074] Call the pre-stored initial Gaussian noise seed As a latent feature representation .

[0075] In this embodiment, the evolutionary trajectory is defined over a continuous interval. Above, among which, Corresponding to the initial Gaussian noise distribution, Distribution of potential watermark features corresponding to the target.

[0076] Set the total number of iterations to .

[0077] Calculate the single-step evolution increment based on the linear discretization principle. .

[0078] Based on the linear interpolation formula, the representation of the suspected latent features is established. Pointing to the latent feature representation as the noise endpoint Deterministic observation path:

[0079] Constructing discrete sequences ,in, , ; The system samples uniformly over time along the aforementioned observation path to obtain the features of the current sampling point corresponding to the time step. .

[0080] This will include deterministic observation paths. Corresponding time embedding and the conditional feature vector to be verified Input to a controlled denoising network with frozen weights, such as Figure 3 As shown.

[0081] Unlike the generation phase, the input for each iteration is no longer updated by the ODE solver and then passed back to the first transformer block.

[0082] Instead, it inputs the features of the corresponding sampling points based on the current discrete time step. Time embedding and the conditional feature vector to be verified .

[0083] The network predicts and outputs the instantaneous velocity vector at that point. .

[0084] The system extracts the residual watermark dynamic bias by comparing the controlled velocity vector predicted by the network with the linear velocity of the observed path itself.

[0085] Predicted instantaneous velocity vector Difference from the baseline path speed :

[0086] For watermarked images, due to image features and There is a coupling relationship during generation. A pointer will be generated at each sampling point. The systematic offset component of the projection direction.

[0087] For watermark-free images, due to image features and Irrelevant It manifests as disordered, random fluctuations.

[0088] Constructing copyright fingerprint feature vectors :

[0089] For spatial dimensions Copyright fingerprint feature vector Perform global pooling to obtain a detection vector with the same dimensions as the watermark vector. .

[0090] Obtain the preset cosine similarity threshold

[0091] Calculate the detection vector With watermark vector cosine similarity :

[0092] when At that time, it was determined that the suspected image contained the watermark information; when At that time, the copyright of the image was excluded.

[0093] Furthermore, as an implementation scheme, Figure 4 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.

[0094] like Figure 4As shown, the computer system may include: a processor 1001, such as a CPU; a memory 1005; a user interface 1003; a network interface 1004; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0095] Those skilled in the art will understand that Figure 4 The computer system architecture shown does not constitute a limitation on the computer system and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0096] like Figure 4 As shown, the memory 1005, as a storage medium, may include an operating system, a network communication module, a user interface module, and computer programs. The operating system is a program that manages and controls the hardware and software resources of the computer system, as well as the operation of the computer programs and other software or programs.

[0097] exist Figure 4 In the computer system shown, the user interface 1003 is mainly used to connect to the terminal and communicate data with the terminal; the network interface 1004 is mainly used to communicate data with the backend server; and the processor 1001 can be used to call the computer program stored in the memory 1005.

[0098] In this embodiment, the computer system includes: a memory 1005, a processor 1001, and a computer program stored in the memory and executable on the processor, wherein: When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: In the watermark generation stage, the watermark information vector is introduced into the construction process of the deterministic evolution trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information. The deterministic evolution trajectory represents the process of the image to be embedded with watermark evolving towards a noise distribution. In the watermark detection stage, Gaussian noise seeds are obtained in the image to be detected. A deterministic observation path corresponding to the image to be detected is determined based on the Gaussian noise seeds. An instantaneous velocity bias is predicted along the deterministic observation path and a copyright fingerprint feature vector is generated. Watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seeds.

[0099] Furthermore, those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in a computer system to implement the process steps of the embodiments of the above methods.

[0100] Therefore, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the trajectory perturbation image watermarking method based on the diffusion model as described in the above embodiments.

[0101] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0102] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.

[0103] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0104] This application 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] 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.

[0106] 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.

[0107] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0108] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A trajectory perturbation image watermarking method based on a diffusion model, characterized in that, The method includes the following steps: In the watermark generation stage, the watermark information vector is introduced into the construction process of the deterministic evolution trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information. The deterministic evolution trajectory represents the process of the image to be embedded with watermark evolving towards a noise distribution. In the watermark detection stage, Gaussian noise seeds are obtained in the image to be detected. A deterministic observation path corresponding to the image to be detected is determined based on the Gaussian noise seeds. An instantaneous velocity bias is predicted along the deterministic observation path and a copyright fingerprint feature vector is generated. Watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seeds.

2. The trajectory perturbation image watermarking method based on the diffusion model as described in claim 1, characterized in that, The watermark information vector is introduced into the construction process of deterministic evolutionary trajectory to generate a controllable bias through real-time trajectory perturbation, forming an endogenous watermark image containing watermark information, specifically including: S10, map the preset watermark sequence into an original feature vector that satisfies the preset spatial dimension of the diffusion model, and modulate the original feature vector into a conditional feature vector that satisfies the image evolution trajectory condition through a nonlinear activation function. Furthermore, the image to be embedded with the watermark is mapped to the latent space to generate initial image features, and the initial image features are linearly interpolated with standard Gaussian noise according to a preset sampling time distribution to construct a deterministic evolution trajectory from the target image distribution to the initial noise distribution; S20, based on the conditional feature vector, the instantaneous velocity vector of the deterministic evolution trajectory is perturbed, and the initial image features are evolved along the deterministic evolution trajectory that has been deflected after the perturbation, so that the watermark sequence can achieve synchronous endogenous coupling during the image generation process, and image features carrying endogenous watermarks are obtained. S30, the image features carrying the endogenous watermark are mapped to the pixel domain to generate the endogenous watermark image.

3. The trajectory perturbation image watermarking method based on the diffusion model as described in claim 2, characterized in that, In step S20, the instantaneous velocity vector in the deterministic evolution trajectory is perturbed according to the conditional feature vector, and the initial image features are evolved along the deterministic evolution trajectory after the perturbation, specifically including: S21, use the sine-cosine coding function to process the current sampling time step. Perform high-dimensional mapping to obtain time-step embedding vectors Embed the time step into the vector With the conditional feature vector Feature fusion is performed, and an adaptive transformation matrix is ​​used. With the corresponding bias term Dynamic response parameters for fitting the denoised trajectory: ； In the formula, and These are the scaling factor and offset factor used to adjust the evolution trajectory, respectively; S22, the dynamic response parameters are applied between the self-attention operator and the feedforward network of the Transformer block to form a controlled vector field function dynamically guided by the watermark. ; S23, the controlled vector field function is subjected to inverse integral evolution from the time step direction through an ordinary differential equation solver; S24, During the integration process, based on the current time step With conditional eigenvectors By fitting the bias of the instantaneous velocity vector in real time, the latent features are driven to evolve along a controlled trajectory, thereby achieving endogenous coupling of the watermark signal during image generation.

4. The trajectory perturbation image watermarking method based on the diffusion model as described in claim 3, characterized in that, Step S23 includes: The feature evolution at each sampling step follows the following linear integral operator: ； In the formula, For single-step evolution increments, Used to characterize the current feature Conditional eigenvectors and time step Evolution rate under common constraints.

5. The trajectory perturbation image watermarking method based on the diffusion model as described in claim 1, characterized in that, Predicting instantaneous velocity bias and generating copyright fingerprint feature vectors along the deterministic observation path specifically includes: S40, invoke the pre-trained encoder to map the image to be detected to the latent space and obtain its corresponding suspected latent feature representation. ; S50, invoke the pre-stored initial Gaussian noise seed, and establish the representation of the suspected latent features according to the linear interpolation formula. Pointing to the latent feature representation as the noise endpoint Deterministic observation path : ； In the formula, Total evolution time For time steps; S60 will define the characteristics of sampling points along the deterministic observation path. The temporal embedding vector corresponding to the features of the sampling points and conditional eigenvectors The input is fed into a controlled denoising network with frozen weights. The instantaneous velocity vector output by the controlled denoising network is compared with the linear velocity vector of the deterministic observation path itself to extract the residual watermark dynamic bias component. S70 performs path integral accumulation on the bias components extracted from each sampling point in the time series, and then performs global pooling to construct a copyright fingerprint feature vector that is consistent with the preset watermark control feature dimension.

6. The trajectory perturbation image watermarking method based on the diffusion model as described in claim 5, characterized in that, The expression for the copyright fingerprint feature vector is: ； In the formula, This represents the total number of discrete sampling steps; Represents the instantaneous velocity vector; For single-step evolution increments; Represents latent features; Indicates potential characteristics of a suspect; This indicates the total evolution time.

7. The trajectory perturbation image watermarking method based on a diffusion model as described in any one of claims 1 to 6, characterized in that, Watermark detection is performed based on the similarity between the copyright fingerprint feature vector and the preset watermark control features associated with the Gaussian noise seed, specifically including: Calculate the cosine similarity between the fingerprint feature vector and the preset watermark control feature; If the cosine similarity is greater than a preset threshold, then the image to be detected is determined to contain endogenous watermark trajectory features. Otherwise, it is determined that it is not included.

8. An application of the trajectory perturbation image watermarking method based on the diffusion model as described in any one of claims 1 to 7, characterized in that, It is used for image watermark generation and / or image watermark detection.

9. A computer system, characterized in that, The computer system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the trajectory perturbation image watermarking method based on the diffusion model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the trajectory perturbation image watermarking method based on the diffusion model as described in any one of claims 1 to 7.

Images