Method and system for generating multiple watermarks by diffusion based on latent variable disturbance and noise scheduling
By introducing a multi-channel latent variable watermark embedding and adaptive noise scheduling module into the latent space of the diffusion model, the problems of watermark embedding capacity, adaptability and robustness in the diffusion model are solved, and efficient and stable embedding and recognition of multi-bit watermarks are achieved.
Patent Information
- Application Number
- CN202610138283.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-31
- Publication Date
- 2026-05-12
AI Technical Summary
Existing digital watermarking methods based on diffusion models suffer from problems such as limited single watermark embedding capacity, lack of adaptability in fixed perturbation strategies, lack of channel-level embedding control, and difficulty in balancing robustness and reversibility.
By introducing a multi-channel latent variable watermark embedding module and an adaptive noise scheduling module into the latent space of the diffusion model, distributed parallel embedding of multi-bit watermarks is achieved, and the embedding amplitude and time position are dynamically adjusted according to the number of diffusion steps and the signal-to-noise ratio.
It significantly improves information capacity and robustness, maintains high image fidelity, supports parallel embedding and recognition of multiple watermarks, has good visual quality and generation stability, and is suitable for various diffusion generation models.
Smart Images
Figure CN122023099A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of digital watermarking and generative artificial intelligence, and particularly relates to a diffusion-based multi-watermark generation method and system based on latent variable perturbation and noise scheduling. Background Technology
[0002] With the widespread application of diffusion models in image generation, editing, and cross-modal generation, the issues of authenticity and copyright protection of generated content have become increasingly prominent. Traditional digital watermarking methods are mainly based on spatial or frequency domain embedding strategies, such as DCT, DWT, or DCT-SVD. These methods have limited effectiveness against compression, regeneration, or deepfake attacks. In recent years, researchers have begun to explore watermark embedding using the latent space characteristics of diffusion models. By encoding implicit information in latent variables or noise space, they have achieved more covert and robust watermarking schemes.
[0003] However, existing watermarking research based on diffusion models mainly suffers from the following problems:
[0004] (1) Single watermark embedding capacity limitation: Most methods only support single watermark information embedding, making it difficult to transmit multiple bits or multiple identity information simultaneously in the same generation process;
[0005] (2) Fixed perturbation strategy lacks adaptability: Traditional diffusion watermarking often embeds information in a fixed noise intensity or diffusion step, and cannot dynamically adjust the embedding intensity according to the diffusion stage, which can easily lead to generation distortion or watermark failure.
[0006] (3) Lack of channel-level embedding control: Existing methods usually superimpose perturbations on the overall latent variables, but do not make full use of the semantic distribution characteristics of the multi-channel latent space of the diffusion model, resulting in interference between different information and affecting the extraction accuracy;
[0007] (4) Robustness and reversibility are difficult to balance: excessively strong embedding perturbation will destroy the generation quality, while excessively weak perturbation will make it difficult to extract effective watermarks after the attack. Summary of the Invention
[0008] To address the aforementioned issues, this invention proposes a diffusion-based multi-watermark generation method and system based on latent variable perturbation and noise scheduling. By introducing a multi-channel latent variable watermark embedding module into the latent space of the diffusion model, distributed parallel embedding of multi-bit watermarks is achieved. Simultaneously, an adaptive noise scheduling module is combined to dynamically adjust the embedding amplitude and time position according to the diffusion steps and signal-to-noise ratio, thereby significantly enhancing the robustness of the watermark while maintaining high image fidelity.
[0009] To achieve the above objectives, this invention provides a diffusion-based multi-watermark generation method and system based on latent variable perturbation and noise scheduling.
[0010] One of the methods for generating multiple watermarks based on latent variable perturbation and noise scheduling includes:
[0011] Redundancy encoding is performed on the input multi-bit watermark information to obtain the watermark encoding sequence;
[0012] Based on the semantic characteristics of the watermarked encoding sequence and the latent variable channels, the watermarked encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain watermarked latent variable channels.
[0013] The noise scheduling parameters are dynamically adjusted based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal.
[0014] According to the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain a disturbed latent variable.
[0015] Perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image;
[0016] Latent variable inversion is performed on the watermarked generated image to obtain the recovered latent variable features;
[0017] Based on the recovered latent variable features, the perturbation offset is detected and decoded to extract the watermark information.
[0018] Preferably, the process of performing redundant encoding on the input multi-bit watermark information to obtain the watermark encoding sequence includes:
[0019] Error correction coding or hash expansion is performed on the original watermark bitstream to obtain an encoded sequence with error recovery capability;
[0020] The watermark encoding sequence is obtained by performing sparse matrix mapping on the encoded sequence.
[0021] Preferably, the process of obtaining the watermarked latent variable channel includes:
[0022] Channel priorities are calculated based on the importance indices of each latent variable channel to obtain channel selection decisions; wherein, the importance indices include gradient change rate and feature entropy value;
[0023] Based on the channel selection decision, different bits of the watermarked encoding sequence are allocated to the corresponding channels to obtain the watermarked latent variable channels.
[0024] Preferably, the process of obtaining the adaptive disturbance intensity control signal includes:
[0025] The embedding intensity adjustment coefficient is calculated based on the signal-to-noise ratio level at the current diffusion time step to obtain the first control component;
[0026] The feedback correction parameters are calculated based on the error curve during the model training phase or the recovery accuracy under attack simulation to obtain the second control component;
[0027] The adaptive disturbance intensity control signal is synthesized based on the first control component and the second control component.
[0028] Preferably, the process of obtaining the perturbated latent variables includes:
[0029] The perturbation direction is determined based on the watermark bit information, and the directional perturbation vector is obtained.
[0030] The directional disturbance vector is scaled and adjusted according to the adaptive disturbance intensity control signal to obtain a disturbance component with controllable intensity.
[0031] The perturbation component with controllable intensity is injected into the watermarked latent variable channel, and the latent variable distribution is kept balanced by regularization constraints to obtain the perturbed latent variable.
[0032] Preferably, the process of obtaining the watermarked generated image includes:
[0033] The perturbated latent variables are denoised step by step, and latent variable consistency constraints are applied at each time step to obtain a constrained latent variable sequence.
[0034] The image is generated by a decoder based on the constrained latent variable sequence, and then post-processed to correct artifacts, thus obtaining the watermarked generated image.
[0035] Preferably, the process of obtaining the characteristics of the recovered latent variables includes:
[0036] The generated image containing the watermark is mapped to the latent variable space by an encoder to obtain a preliminary latent variable estimate;
[0037] Based on the preliminary latent variable estimation, the latent variable trajectory is reconstructed using a noise matching algorithm to obtain the recovered latent variable characteristics.
[0038] Preferably, the process of extracting and obtaining watermark information includes:
[0039] The disturbance offset is calculated by comparing the recovered latent variable characteristics with the baseline distribution to obtain the disturbance estimate.
[0040] Based on the disturbance estimate, the binary bit stream is reconstructed through amplitude threshold determination or direction identification to obtain the extracted watermark information;
[0041] When multiple watermarks exist, they are extracted in groups based on the channel index or identification code to obtain separate groups of watermark bit streams.
[0042] Preferably, the method further includes:
[0043] By constraining the perturbation directions of different watermarks based on the orthogonal partitioning of latent variable channels, latent variables of multi-watermark orthogonal embedding are obtained.
[0044] By optimizing the upper limit of the perturbation amplitude based on adversarial training and robustness constraints, watermark embedding parameters resistant to attacks are obtained.
[0045] This invention also provides a diffusion-generated multi-watermark system based on latent variable perturbation and noise scheduling, comprising:
[0046] A multi-channel latent variable watermarking embedding module is used to perform redundant encoding based on the input multi-bit watermark information to obtain a watermark encoding sequence; based on the semantic characteristics of the watermark encoding sequence and the latent variable channels, the watermark encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain watermarked latent variable channels.
[0047] An adaptive noise scheduling module is used to dynamically adjust noise scheduling parameters based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal; based on the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain a disturbed latent variable.
[0048] The diffusion generation and backsampling module is used to perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image; and to perform latent variable inversion based on the watermarked generated image to obtain the recovered latent variable features.
[0049] The watermark recognition module is used to detect and decode the disturbance offset based on the recovered latent variable features, and extract the watermark information.
[0050] Compared with the prior art, the present invention has the following advantages and technical effects:
[0051] (1) Significantly improved information capacity and robustness: This invention overcomes the bottleneck of limited capacity in traditional single-channel or image domain watermark embedding by introducing a multi-channel latent variable perturbation mechanism. The latent variable space of the diffusion model has high-dimensionality and redundant feature distribution characteristics, and different channels are semantically weakly correlated, which makes it possible to embed watermark information in multiple channels simultaneously. By performing gradient analysis and energy constraints on channel importance, this method achieves parallel embedding of multi-bit watermarks without destroying the semantic structure of the image, and the information capacity is improved by about 3-5 times compared with the traditional diffusion watermarking method.
[0052] Furthermore, the adaptive noise scheduling module dynamically adjusts the embedding strength at different stages of diffusion and back-diffusion, ensuring the watermark signal remains recoverable during the anti-noise propagation path. Even when the model suffers distortions such as JPEG compression, random cropping, Gaussian noise contamination, or re-diffusion attacks, the watermark can still be extracted with high accuracy. This latent space robust embedding method fundamentally changes the fragile structure of traditional watermarks, giving the watermark a "regenerative" characteristic at the generation and distribution level; that is, even if the image is regenerated through diffusion, its underlying watermark pattern can still be recovered.
[0053] (2) Significantly optimized visual quality and generation stability: Traditional image domain watermarking methods often cause visible distortion due to direct perturbation of pixels. However, this invention embeds watermarks at the latent variable level, and protects the quality of image generation through perturbation constraints and consistency regularization. By introducing latent variable norm constraints and a perceptual loss function, the system ensures that the embedded perturbation is smoothly distributed in the latent space, and the generated image maintains a natural and perceptually indistinguishable appearance.
[0054] An adaptive noise scheduling module further balances embedding strength and generation stability. Suppressing perturbation strength during high signal-to-noise ratio (SNR) stages prevents early feature corruption, while gradually enhancing embedding during low SNR stages improves signal retention, thus achieving smooth embedding propagation throughout the diffusion reverse process. The optimized scheduling strategy maintains the optimal balance between generation quality and watermark strength without increasing model complexity.
[0055] (3) Improved multi-watermark parallelism and identity recognition mechanism: This invention supports independent embedding and recognition mechanisms for multiple watermarks, providing a solution for multi-party coexistence and source tracing in complex copyright environments. Through the orthogonal partitioning of latent variable channels and perturbation direction constraints, different watermarks can achieve spatial isolation and information orthogonality during embedding, avoiding the "interference and overlay" problems common in traditional multi-watermarking schemes.
[0056] During the extraction phase, the system can quickly extract watermarks in groups based on the channel index table or embedded identity codes, enabling parallel recognition of multiple watermarks. The watermark recognition module incorporates a lightweight matching network to perform similarity matching between the extracted bitstream and watermark patterns in the registration database, and combines this with a confidence threshold to determine watermark attribution. This structure can simultaneously support three application modes: "watermark verification," "multi-source identity differentiation," and "copyright tracing."
[0057] The biggest advantage of this mechanism is that it can embed multiple identity markers simultaneously during a single generation process, such as the authentication watermarks of content creators, distribution platforms, and regulatory agencies, without interference between them. This feature has broad practical value in the future AIGC ecosystem.
[0058] (4) Enhanced Application Value and System Extensibility: The design of this invention is not limited to copyright protection of static images. Its methodological structure can be extended to video generation, 3D content generation, and even cross-modal diffusion models. Since the embedding process is based on latent variable perturbation rather than fixed pixel operation, the system can be seamlessly migrated to different types of diffusion generators, such as StableDiffusion, VideoDiffusion, and 3D Gaussian Splatting Diffusion.
[0059] At the application level, this method can achieve the following functions:
[0060] Copyright protection and anti-counterfeiting traceability: By embedding a unique identity watermark, the source of AI-generated images is authenticated and tamper-proofed and traced.
[0061] AIGC content tracing and regulation: Providing verifiable "invisible identity codes" for generated content helps prevent its misuse;
[0062] Model security and ownership identification: Manufacturer identifiers can be embedded in the weights or intermediate latent variables of the diffusion model to prevent model theft and reverse distribution;
[0063] Scalable implicit communication: Multi-level information can be embedded in generated content through multiple watermarking channels to achieve implicit data transmission or joint authentication.
[0064] From a system architecture perspective, each module of this invention possesses a high degree of independence and substitutability. The latent variable watermark embedding module can adapt to different model architectures (such as U-Net, DiT, Transformer-based Diffuser), the adaptive noise scheduling strategy can work in conjunction with any noise planning function (linear, cosine, DDIM, etc.), and the watermark recognition module can select traditional matching or neural discriminant methods according to the application scenario, demonstrating good engineering feasibility and scalability.
[0065] In summary, this invention has significant advantages in terms of technological innovation, robustness, scalability, and application value. It provides a new technical path for the secure generation and trusted identification of diffusion models and lays the foundation for content trustworthiness and copyright compliance in the AIGC field. Attached Figure Description
[0066] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0067] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0068] Figure 2 This is a flowchart illustrating the encoding and embedding mapping process for generating watermark information according to an embodiment of the present invention.
[0069] Figure 3 This is a flowchart illustrating the dynamic noise scheduling and diffusion generation process according to an embodiment of the present invention. Detailed Implementation
[0070] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0071] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0072] Example 1
[0073] like Figure 1 As shown, this embodiment provides a diffusion-based multi-watermark generation method based on latent variable perturbation and noise scheduling, including:
[0074] Redundancy encoding is performed on the input multi-bit watermark information to obtain the watermark encoding sequence;
[0075] Based on the semantic characteristics of the watermarked encoding sequence and the latent variable channels, the watermarked encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain the watermarked encoding latent variable channels.
[0076] The noise scheduling parameters are dynamically adjusted based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal.
[0077] Based on the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain the disturbed latent variable.
[0078] Perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image;
[0079] Latent variable inversion is performed on the watermarked generated image to obtain the recovered latent variable features;
[0080] The perturbation offset is detected and decoded based on the recovered latent variable features, and the watermark information is extracted.
[0081] Furthermore, the process of performing redundant encoding on the input multi-bit watermark information to obtain the watermark encoded sequence includes:
[0082] Error correction coding or hash expansion is performed on the original watermark bitstream to obtain an encoded sequence with error recovery capability;
[0083] The watermarked sequence is obtained by performing sparse matrix mapping on the encoded sequence.
[0084] Furthermore, such as Figure 2 As shown, the process of obtaining a watermarked latent variable channel includes:
[0085] Channel priority is calculated based on the importance indices of each channel of the latent variable to obtain channel selection decisions; among which, the importance indices include gradient change rate and feature entropy value;
[0086] Based on the channel selection decision, different bits of the watermarked encoding sequence are allocated to the corresponding channels to obtain watermarked latent variable channels.
[0087] Furthermore, in this embodiment, redundant encoding and channel mapping are implemented through a multi-channel latent variable watermarking embedding module. This module is responsible for embedding multi-bit watermark information into the latent variable space of the diffusion model. During the encoding stage, the diffusion model converts the input image or random noise into a latent variable representation. The latent variables are typically composed of multiple channels, each containing semantic features at different levels. This module utilizes this structural characteristic to allocate and embed the watermark information into multiple latent variable channels.
[0088] The embedding process includes the following steps:
[0089] First, the watermark encoding stage. The system first performs redundant encoding on the original watermark information, such as using error correction coding or hash expansion, so that it can still be recovered after being subjected to channel disturbances.
[0090] Second, the channel allocation stage. The system dynamically selects the embedding position based on the importance indicators of each channel of the latent variable (such as gradient change rate, feature entropy value, etc.), and allocates the information to relatively stable or low-sensitivity channels.
[0091] Third, the perturbation injection stage. The system applies a perturbation with controllable amplitude to the selected channel. The direction of the perturbation corresponds to the watermark bit information, and the intensity of the perturbation is controlled by an adjustment coefficient, thereby achieving implicit information injection without destroying the latent variable structure.
[0092] Fourth, the perturbation constraint stage. To avoid the embedding process from significantly affecting image quality, the system sets an upper limit on the perturbation intensity and maintains the overall balance of the latent variable distribution through regularization constraints.
[0093] Through the above steps, this module can achieve high-capacity watermark embedding at the latent space level, with not only high embedding accuracy but also strong recoverability during subsequent diffusion generation.
[0094] Furthermore, such as Figure 3 As shown, the process of obtaining the adaptive disturbance intensity control signal includes:
[0095] The embedding intensity adjustment coefficient is calculated based on the signal-to-noise ratio level at the current diffusion time step to obtain the first control component;
[0096] The feedback correction parameters are calculated based on the error curve during the model training phase or the recovery accuracy under attack simulation to obtain the second control component;
[0097] An adaptive disturbance intensity control signal is synthesized based on the first control component and the second control component.
[0098] Furthermore, the process of obtaining the perturbed latent variables includes:
[0099] The perturbation direction is determined based on the watermark bit information, and the directional perturbation vector is obtained.
[0100] The amplitude of the directional disturbance vector is adjusted by scaling the adaptive disturbance intensity control signal to obtain a disturbance component with controllable intensity.
[0101] By injecting a perturbation component with controllable intensity into the watermarked latent variable channel and maintaining the distribution balance of the latent variables through regularization constraints, the perturbed latent variables are obtained.
[0102] Furthermore, the dynamic adjustment of noise scheduling parameters in this embodiment is achieved through an adaptive noise scheduling module. This module dynamically adjusts the embedding strength and position of the watermark based on the number of diffusion steps and the signal-to-noise ratio (SNR) to ensure the stability of the watermark signal transmission during diffusion and reverse diffusion. In the diffusion model, the image undergoes a forward diffusion process from clear to completely noisy, and then is gradually restored to a true image through reverse diffusion. This module automatically adjusts the watermark embedding strategy based on the noise level at each diffusion stage. When the SNR is high, the system reduces the embedding strength to avoid affecting early features; when the SNR is low, the system gradually strengthens the embedding to ensure stable transmission of the watermark signal during denoising and restoration.
[0103] The adaptive scheduling module also features feedback adjustment capabilities, automatically correcting embedding parameters based on the model's error curve during training or its recovery accuracy under attack simulations, thus achieving dynamic optimization. This mechanism ensures high watermark fidelity and generation consistency across different diffusion strategies (such as linear, cosine, and nonlinear noise scheduling).
[0104] Furthermore, the process of obtaining the watermarked generated image includes:
[0105] The perturbated latent variables are denoised step by step, and consistency constraints are applied to the latent variables at each time step to obtain the constrained latent variable sequence.
[0106] The image is generated by a decoder based on the constrained latent variable sequence, and then post-processed to correct artifacts, resulting in a watermarked generated image.
[0107] Furthermore, the process of obtaining the characteristics of the recovered latent variables includes:
[0108] The generated image with watermark is mapped to the latent variable space by an encoder to obtain a preliminary estimate of the latent variables;
[0109] Based on the preliminary latent variable estimates, the latent variable trajectories are reconstructed using a noise matching algorithm to obtain the recovered latent variable characteristics.
[0110] Furthermore, in this embodiment, backdiffusion sampling and image generation, as well as latent variable inversion and perturbation detection, are implemented through a diffusion generation and backsampling module. This module completes the generation process from latent variables to a visible image and performs inverse mapping when necessary to recover the latent variable structure. The system first gradually adds noise during the forward diffusion stage to generate a multi-layer latent variable sequence, and then gradually removes the noise using a pre-trained denoising model during the backdiffusion stage, thereby generating a watermarked image.
[0111] During the generation process, the system introduces latent variable consistency constraints to preserve embedding perturbations during the generation stage without significantly affecting the image content. This module also supports re-diffusion operations, i.e., performing a diffusion-anti-diffusion process again on the watermarked image to verify the robustness of the watermark under multiple generation rounds. This module maintains the naturalness of the image while ensuring that the watermark information in the latent space is not destroyed.
[0112] Furthermore, the process of extracting the watermark information includes:
[0113] The disturbance offset is calculated by comparing the recovered latent variable characteristics with the baseline distribution, and the disturbance estimate is obtained.
[0114] Based on the disturbance estimate, the binary bit stream is reconstructed through amplitude threshold determination or direction identification to obtain the extracted watermark information;
[0115] When multiple watermarks exist, they are extracted in groups based on the channel index or identification code to obtain separate groups of watermark bit streams.
[0116] Furthermore, the method also includes:
[0117] By constraining the perturbation directions of different watermarks based on the orthogonal partitioning of latent variable channels, latent variables of multi-watermark orthogonal embedding are obtained.
[0118] By optimizing the upper limit of the perturbation amplitude based on adversarial training and robustness constraints, watermark embedding parameters resistant to attacks are obtained.
[0119] Furthermore, in this embodiment, the extraction of watermark information and the performance of similarity matching and identity determination are achieved through a watermark recognition module. This module is used to extract latent variable features from watermarked images and reconstruct the watermark information. The extraction process includes the following steps:
[0120] Encoding and restoration. Using the same encoder structure as the embedding stage, the input image is remapped into the latent variable space;
[0121] Perturbation detection. The system estimates the perturbation shifts present in the latent variables by comparing their characteristics with the baseline distribution.
[0122] Bit decoding. The system restores the latent variable changes to a binary bitstream based on the direction or amplitude threshold of the perturbation. Multi-watermark separation: If multiple watermarks exist, they are extracted in groups based on channel index or identification code to distinguish embedded information from different sources.
[0123] This module can extract watermarks without accessing the weight parameters of the original diffusion model, and has high independence and applicability.
[0124] Simultaneously, this module is also used to determine whether the extracted watermark bitstream matches the target identity, completing copyright verification or source identification. The identification process includes three stages: watermark matching, threshold determination, and identity differentiation.
[0125] First, the system uses a lightweight recognition network to calculate the similarity between the extracted bitstream and watermark templates in the registration database. Then, it determines whether a match is successful based on a set confidence threshold. Finally, if multiple watermarks coexist, the system can distinguish different embedding sources based on a preset channel mapping table or identity identifier, thereby achieving multi-entity identity recognition and tracing.
[0126] This module supports parallel recognition and real-time determination, and can be expanded to application scenarios such as copyright verification, model ownership detection, and joint certification by multiple institutions.
[0127] This embodiment applies a light perturbation to each channel in the latent variable layer, with different channels carrying different watermark bits, and reduces inter-channel interference through an orthogonalization strategy. The noise scheduling module dynamically adjusts the perturbation intensity according to the noise weight function β(t) or σ(t), enabling coordinated optimization of the watermark embedding process and the diffusion generation stage. Furthermore, multi-watermark information can be extracted and identified through back-diffusion sampling or intermediate layer feature recovery. This method can accurately identify multi-bit watermark information even under JPEG compression, random cropping, re-diffusion, and resampling attacks, and the generated image quality is close to that of the original model.
[0128] This embodiment embeds multi-bit watermark information into the diffusion generation process by perturbing multi-channel latent variables in the latent space of the diffusion model. First, the input watermark sequence is encoded and mapped to latent variable channels, achieving parallel embedding of multiple watermarks. Second, noise scheduling parameters are dynamically adjusted based on the number of diffusion steps and the signal-to-noise ratio, employing an adaptive perturbation strategy to enhance the robustness and reversibility of the embedding. Finally, watermark information can be extracted and identified during the backsampling or generation stage. This method offers advantages such as high capacity, low distortion, and strong robustness, resisting various attacks including compression, pruning, and re-diffusion. It is suitable for applications such as diffusion model content tracing, copyright authentication, and anti-counterfeiting of generated content.
[0129] This embodiment improves overall system performance through a joint optimization strategy. The optimization objectives include three parts: the reconstruction accuracy of the generated image, the consistency of watermark embedding and extraction, and the system's resistance to attacks. Perceptual loss and adversarial noise simulation are introduced during the training phase, enabling the system to maintain high recognition accuracy under image compression, cropping, resampling, and re-diffusion attacks.
[0130] Furthermore, the multi-watermark orthogonal embedding mechanism proposed in this embodiment ensures that different watermark signals are independent of each other, avoiding cross-interference. Through adversarial training and robustness constraints, the system can maintain stable performance in complex generation environments, ensuring that the watermarks are both concealed and recoverable.
[0131] In summary, this embodiment proposes a method that integrates latent variable perturbation mechanism and adaptive noise scheduling to achieve high-capacity, multi-watermark, and robust embedding and recognition while ensuring generation quality. This method has broad practical value and innovative significance.
[0132] As a preferred implementation, this embodiment provides a diffusion-based multi-watermark generation method based on latent variable perturbation and noise scheduling. This method introduces a multi-level perturbation strategy and a dynamic noise scheduling mechanism into the latent space of the diffusion model, achieving adaptive embedding, robust extraction, and enhanced traceability of multi-layer watermarks. This method combines implicit latent variable control with explicit frequency domain perturbation, enabling the generated image to possess multiple anti-counterfeiting and identification capabilities while maintaining quality. The implementation process of this embodiment will be described in detail below. (Refer to...) Figure 1 The implementation steps of this embodiment are as follows:
[0133] I. Initialization and Input Preparation of the Diffusion Model;
[0134] like Figure 1 As shown, this embodiment first establishes a basic generation framework based on a pre-trained diffusion generation model. The user inputs text prompts and multiple sets of watermark data through the system interface. The system guides image semantic generation based on the input text information, and simultaneously configures latent variable perturbation parameters based on the input multi-channel watermark information.
[0135] (a) Input Materials: Users input text prompts, main watermark information, and auxiliary watermark information. The main watermark represents the core copyright identifier or traceability information, while the auxiliary watermark can carry additional authentication content or classification marks. The system encodes the input text, generates text feature vectors, and assigns embedding weights to the watermark data.
[0136] (ii) Model Loading: The system loads a pre-trained diffusion generation model, including a forward diffusion module, a backward generation module, and a latent variable mapping network. The model parameters are dynamically managed by the scheduler to support adaptive adjustments to subsequent noise intensity and perturbation amplitude.
[0137] (III) Latent variable initialization: Before the input image is generated, the system samples latent variables from a Gaussian distribution and establishes a perturbation channel related to the watermark information through a latent space encoder. This latent variable will serve as the basic carrier for watermark embedding and diffusion inversion.
[0138] II. Watermark Information Encoding and Embedding Mapping;
[0139] The goal of this stage is to transform multiple sets of watermark information into embeddable latent variable perturbation representations, so that they can be naturally integrated into the latent space structure during the generation process as noise diffuses.
[0140] (i) Multi-channel watermarking encoding: The system encodes multiple sets of watermark information into independent binary sequences and performs error-checking encoding and sparse matrix mapping on them. The main watermark adopts a high-stability encoding scheme, while the auxiliary watermark adopts a scalable lightweight encoding scheme. The two form a differentiated embedding through a position weight distribution strategy.
[0141] (ii) Perturbation Vector Generation: The encoded watermark sequence is transformed into a continuous perturbation vector through a perturbation mapping network, and the perturbation intensity is dynamically controlled by a noise scheduler. The system maps the watermark vector to the mid-to-high frequency region of the latent variable distribution to ensure that the watermark information can be preserved during the inversion process of model sampling.
[0142] (iii) Latent variable fusion: The system weights and fuses the perturbation vector with the original latent variables to form a mixed latent variable. The perturbation weight corresponding to the main watermark is higher to ensure the stability of source tracing; the perturbation of the auxiliary watermark is lower to reduce the impact on the image structure.
[0143] III. Dynamic noise scheduling and diffusion generation;
[0144] This stage is the core of diffusion generation and watermark embedding, and the disturbance injection process is precisely controlled through a noise scheduling mechanism.
[0145] (i) Noise intensity scheduling: The system employs a dynamic noise scheduling algorithm during the diffusion inversion process, adjusting the noise injection ratio in real time based on the number of time steps and image complexity. The noise scheduling follows a smooth decreasing law, making the embedded signal more concealed in the early stage and enhancing robustness in the later stage.
[0146] (ii) Latent variable diffusion: In each back diffusion step, the system performs noise removal and perturbation retention operations. The model predicts the latent variable residuals and corrects the watermark embedding error through a multi-layer neural network. The dynamic response parameters of the perturbation are recorded at each diffusion time step for subsequent source tracing analysis.
[0147] (III) Parallel Control of Multiple Watermarks: The system supports the synchronous diffusion generation of multiple watermarks, and the distribution and embedding depth of different watermark channels are controlled by independent perturbation masks. This mechanism ensures that multiple watermarks do not interfere with each other and can be separated and identified during extraction.
[0148] (iv) Image Output Generation: After diffusion inversion is completed, the system decodes the latent variables to obtain the final image, which contains both implicit latent variable watermarks and explicit extractable digital watermarks. The generated image undergoes quality assessment and artifact correction by the post-processing module to ensure a natural visual effect without obvious embedding traces.
[0149] IV. Watermark Extraction and Verification Process;
[0150] During the image publishing or verification phase, the system recovers the embedded watermark information through inverse diffusion inversion and feature matching.
[0151] (i) Latent variable inversion: The system inputs the image to be verified into the diffusion model inversion module to recover the approximate distribution of latent variables. During the inversion process, the trajectory of latent variables is reconstructed through a noise matching algorithm, thereby recovering the perturbation features in the latent space.
[0152] (II) Watermark Decoding: The system utilizes a watermark extraction network to analyze and reconstruct latent variable disturbance signals. High-frequency disturbances in the main watermark are decoded to recover core traceability information; mid-to-low-frequency disturbances in the auxiliary watermark are recovered as supplementary identification data. The extraction process employs residual correlation measurement and error correction mechanisms to ensure stable decoding.
[0153] (iii) Information Verification: The extracted watermark information will be compared with the information registered in the central database or blockchain. If the match is successful, the authenticity and source of the image are confirmed; if they are inconsistent, the system will generate an anomaly report and mark the potential risk of tampering.
[0154] V. Image traceability and management;
[0155] (i) Source tracking record generation: The system tracks the image's generation path, usage records, and propagation chain based on the unique identifier of the master watermark. All verification data and metadata are synchronously written to the blockchain to achieve tamper-proof distributed records.
[0156] (ii) Multidimensional association analysis: The system can perform association analysis in the database based on the watermark identifier to retrieve images from the same source or series generated by the same model, which helps to determine the legality of the image source.
[0157] (III) Secure Publishing and Access Control: The system manages the publishing permissions of watermarked images through an encryption mechanism, and only authorized users can access the original data. Images published externally still retain latent variable perturbation characteristics, and even if they are compressed, cropped, or re-edited, their authenticity can still be verified through the extraction module.
[0158] The entire system forms a closed-loop mechanism from generation, embedding, release, verification, traceability to tracking. The model can dynamically adjust parameters at each stage to achieve scalable multi-watermark embedding and stable extraction performance.
[0159] In this embodiment, the combination of latent variable perturbation and noise scheduling mechanisms significantly improves the watermark's concealment and resistance to attacks. Compared with traditional embedding methods based on the spatial or frequency domains, this method achieves natural fusion within the diffusion generation framework, which not only preserves the high-fidelity characteristics of the generated image but also effectively resists subsequent image tampering, compression, and recoding attacks.
[0160] Example 2
[0161] Based on the same inventive concept, this embodiment also provides a diffusion-generated multi-watermark system based on latent variable perturbation and noise scheduling, including:
[0162] The multi-channel latent variable watermarking embedding module is used to perform redundant encoding based on the input multi-bit watermark information to obtain a watermark encoding sequence; based on the semantic characteristics of the watermark encoding sequence and the latent variable channels, the watermark encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain watermarked latent variable channels.
[0163] An adaptive noise scheduling module is used to dynamically adjust noise scheduling parameters based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal; based on the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain a disturbed latent variable.
[0164] The diffusion generation and backsampling module is used to perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image; and to perform latent variable inversion based on the watermarked generated image to obtain the recovered latent variable features.
[0165] The watermark recognition module is used to detect and decode the disturbance offset based on the recovered latent variable features, and extract the watermark information.
[0166] This embodiment specifically relates to a diffusion model multi-watermark embedding and recognition system that combines latent variable perturbation and adaptive noise scheduling. This technology can be widely used in copyright protection of image-generated content, model security protection, and multi-identity watermark verification systems.
[0167] The diffusion-generated multi-watermarking system based on latent variable perturbation and noise scheduling provided in this embodiment has all the advantages of the diffusion-generated multi-watermarking method based on latent variable perturbation and noise scheduling provided in Embodiment 1.
[0168] Example 3
[0169] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.
[0170] Example 4
[0171] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0172] Example 5
[0173] This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0174] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A diffusion-based multi-watermark generation method based on latent variable perturbation and noise scheduling, characterized in that, include: Redundancy encoding is performed on the input multi-bit watermark information to obtain the watermark encoding sequence; Based on the semantic characteristics of the watermarked encoding sequence and the latent variable channels, the watermarked encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain watermarked latent variable channels. The noise scheduling parameters are dynamically adjusted based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal. According to the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain a disturbed latent variable. Perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image; Latent variable inversion is performed on the watermarked generated image to obtain the recovered latent variable features; Based on the recovered latent variable features, the perturbation offset is detected and decoded to extract the watermark information.
2. The method according to claim 1, characterized in that, The process of performing redundant encoding on the input multi-bit watermark information to obtain the watermark encoded sequence includes: Error correction coding or hash expansion is performed on the original watermark bitstream to obtain an encoded sequence with error recovery capability; The watermark encoding sequence is obtained by performing sparse matrix mapping on the encoded sequence.
3. The method according to claim 1, characterized in that, The process of obtaining watermarked latent variable channels includes: Channel priorities are calculated based on the importance indices of each latent variable channel to obtain channel selection decisions; wherein, the importance indices include gradient change rate and feature entropy value; Based on the channel selection decision, different bits of the watermarked encoding sequence are allocated to the corresponding channels to obtain the watermarked latent variable channels.
4. The method according to claim 1, characterized in that, The process of obtaining the adaptive disturbance intensity control signal includes: The embedding intensity adjustment coefficient is calculated based on the signal-to-noise ratio level at the current diffusion time step to obtain the first control component; The feedback correction parameters are calculated based on the error curve during the model training phase or the recovery accuracy under attack simulation to obtain the second control component; The adaptive disturbance intensity control signal is synthesized based on the first control component and the second control component.
5. The method according to claim 1, characterized in that, The process of obtaining the perturbed latent variables includes: The perturbation direction is determined based on the watermark bit information, and the directional perturbation vector is obtained. The directional disturbance vector is scaled and adjusted according to the adaptive disturbance intensity control signal to obtain a disturbance component with controllable intensity. The perturbation component with controllable intensity is injected into the watermarked latent variable channel, and the latent variable distribution is kept balanced by regularization constraints to obtain the perturbed latent variable.
6. The method according to claim 1, characterized in that, The process of obtaining a watermarked image includes: The perturbated latent variables are denoised step by step, and latent variable consistency constraints are applied at each time step to obtain a constrained latent variable sequence. The image is generated by a decoder based on the constrained latent variable sequence, and then post-processed to correct artifacts, thus obtaining the watermarked generated image.
7. The method according to claim 1, characterized in that, The process of obtaining the characteristics of the recovered latent variables includes: The generated image containing the watermark is mapped to the latent variable space by an encoder to obtain a preliminary latent variable estimate; Based on the preliminary latent variable estimation, the latent variable trajectory is reconstructed using a noise matching algorithm to obtain the recovered latent variable characteristics.
8. The method according to claim 1, characterized in that, The process of extracting watermark information includes: The disturbance offset is calculated by comparing the recovered latent variable characteristics with the baseline distribution to obtain the disturbance estimate. Based on the disturbance estimate, the binary bit stream is reconstructed by amplitude threshold determination or direction identification to obtain the extracted watermark information; When multiple watermarks exist, they are extracted in groups based on the channel index or identification code to obtain separate groups of watermark bit streams.
9. The method according to claim 1, characterized in that, The method further includes: By constraining the perturbation directions of different watermarks based on the orthogonal partitioning of latent variable channels, latent variables of multi-watermark orthogonal embedding are obtained. By optimizing the upper limit of the perturbation amplitude based on adversarial training and robustness constraints, watermark embedding parameters resistant to attacks are obtained.
10. A diffusion-generated multi-watermark system based on latent variable perturbation and noise scheduling, characterized in that, include: The multi-channel latent variable watermarking embedding module is used to perform redundant encoding based on the input multi-bit watermark information to obtain the watermark encoding sequence. Based on the semantic characteristics of the watermarked encoding sequence and the latent variable channels, the watermarked encoding sequence is mapped to multiple channels of the latent variables in the diffusion model to obtain watermarked latent variable channels. An adaptive noise scheduling module is used to dynamically adjust noise scheduling parameters based on the number of diffusion steps and the signal-to-noise ratio to obtain an adaptive disturbance intensity control signal; based on the adaptive disturbance intensity control signal, a disturbance with controllable amplitude is injected into the watermarked latent variable channel to obtain a disturbed latent variable. The diffusion generation and backsampling module is used to perform backdiffusion sampling based on the perturbed latent variables to obtain a watermarked generated image; and to perform latent variable inversion based on the watermarked generated image to obtain the recovered latent variable features. The watermark recognition module is used to detect and decode the disturbance offset based on the recovered latent variable features, and extract the watermark information.