CN-122023099-A - Diffusion generation multi-watermark method and system based on latent variable disturbance and noise scheduling

CN122023099ACN 122023099 ACN122023099 ACN 122023099ACN-122023099-A

Abstract

The invention discloses a diffusion generation multi-watermark method and system based on latent variable disturbance and noise scheduling, which comprises the steps of conducting redundancy coding according to input multi-bit watermark information to obtain a watermark coding sequence, mapping the watermark coding sequence to a plurality of channels of diffusion model latent variables according to semantic characteristics of the watermark coding sequence and the latent variable channels to obtain the latent variable channels, dynamically adjusting noise scheduling parameters according to diffusion step numbers and signal to noise ratios to obtain adaptive disturbance intensity control signals, injecting disturbance with controllable amplitude into the latent variable channels with the watermark coding to obtain latent variables with disturbance, conducting back diffusion sampling according to the latent variables with disturbance to obtain a generated image with the watermark, conducting latent variable inversion according to the generated image with the watermark to obtain recovered latent variable characteristics, detecting disturbance offset according to the recovered latent variable characteristics, decoding, and extracting to obtain watermark information. The invention has the advantages of high capacity, low distortion, strong robustness and the like.

Inventors

ZHU XIAOQIANG
TIAN YE
ZENG DAN

Assignees

上海大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (10)

1. A method for generating multiple watermarks based on diffusion of latent variable disturbance and noise scheduling, comprising: performing redundancy coding according to the input multi-bit watermark information to obtain a watermark coding sequence; According to the semantic characteristics of the watermark coding sequence and the latent variable channels, mapping the watermark coding sequence to a plurality of channels of diffusion model latent variables to obtain the latent variable channels with watermark codes; Dynamically adjusting noise scheduling parameters according to the number of diffusion steps and the signal-to-noise ratio to obtain a self-adaptive disturbance intensity control signal; Injecting disturbance with controllable amplitude into the latent variable channel with watermark coding according to the self-adaptive disturbance intensity control signal to obtain a latent variable with disturbance; Performing back diffusion sampling according to the latent variable with disturbance to obtain a generated image containing the watermark; performing latent variable inversion according to the generated image containing the watermark to obtain recovered latent variable characteristics; And detecting disturbance offset according to the restored latent variable characteristics, decoding, and extracting to obtain watermark information.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, Redundancy coding is carried out according to the input multi-bit watermark information, and the process for obtaining the watermark coding sequence comprises the following steps: Performing error correction coding or hash expansion according to the original watermark bit stream to obtain a coded sequence with error recovery capability; and performing sparse matrix mapping according to the coded sequence to obtain the watermark coding sequence.
3. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining the latent variable channel with the watermark code comprises the following steps: calculating channel priority according to importance indexes of each channel of the latent variable to obtain a channel selection decision, wherein the importance indexes comprise gradient change rate and characteristic entropy; And distributing different bits of the watermark coding sequence to corresponding channels according to the channel selection decision to obtain the latent variable channel with watermark coding.
4. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining the adaptive disturbance intensity control signal includes: Calculating an embedded strength adjustment coefficient according to the signal-to-noise ratio level of the current diffusion time step to obtain a first control component; calculating a feedback correction parameter according to an error curve of a model training stage or recovery accuracy under attack simulation to obtain a second control component; And synthesizing the adaptive disturbance intensity control signal according to the first control component and the second control component.
5. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining the perturbed latent variable includes: Determining a disturbance direction according to watermark bit information to obtain a directional disturbance vector; performing amplitude scaling adjustment on the directional disturbance vector according to the self-adaptive disturbance intensity control signal to obtain a disturbance component with controllable intensity; And injecting the disturbance component with controllable intensity into the latent variable channel with the watermark coding, and keeping the distribution balance of the latent variable through regularization constraint to obtain the latent variable with disturbance.
6. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining a watermarked generated image comprises: step-by-step denoising is carried out on the latent variable with disturbance, and a latent variable consistency constraint is applied to each time step, so that a constrained latent variable sequence is obtained; And generating an image through a decoder according to the constrained latent variable sequence, and correcting the artifact through post-processing to obtain the generated image containing the watermark.
7. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining the recovered latent variable features includes: mapping the generated image containing the watermark to a latent variable space through an encoder to obtain a preliminary latent variable estimation; reconstructing a latent variable track through a noise matching algorithm according to the preliminary latent variable estimation, and obtaining the recovered latent variable characteristics.
8. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of extracting watermark information comprises the following steps: calculating disturbance offset according to the comparison between the restored latent variable characteristics and the reference distribution, and obtaining a disturbance estimated value; according to the disturbance estimated value, the binary bit stream is restored through amplitude threshold judgment or direction recognition, and the extracted watermark information is obtained; when a plurality of watermarks exist, grouping extraction is carried out according to the channel index or the identity identification code, and separated multi-group watermark bit streams are obtained.
9. The method of claim 1, wherein the step of determining the position of the substrate comprises, The method further comprises the steps of: Constraining disturbance directions of different watermarks according to orthogonal division of the latent variable channel to obtain a multi-watermark orthogonal embedded latent variable; Optimizing the upper limit of disturbance amplitude according to the countermeasure training and the robustness constraint, and obtaining the watermark embedding parameter of the attack resistance.
10. A system for generating multiple watermarks based on diffusion of latent variable disturbance and noise scheduling, comprising: The multi-channel latent variable watermark embedding module is used for carrying out redundancy coding according to the input multi-bit watermark information to obtain a watermark coding sequence, and mapping the watermark coding sequence to a plurality of channels of a diffusion model latent variable according to the semantic characteristics of the watermark coding sequence and the latent variable channels to obtain a latent variable channel with watermark coding; The self-adaptive noise scheduling module is used for dynamically adjusting noise scheduling parameters according to the diffusion step number and the signal to noise ratio to obtain a self-adaptive disturbance intensity control signal; The diffusion generation and reverse sampling module is used for executing reverse diffusion sampling according to the disturbance latent variable to obtain a watermark-containing generated image; And the watermark identification module is used for detecting disturbance offset according to the restored latent variable characteristics, decoding and extracting to obtain watermark information.

Description

Diffusion generation multi-watermark method and system based on latent variable disturbance and noise scheduling Technical Field The invention belongs to the field of digital watermarking and generation type artificial intelligence, and particularly relates to a method and a system for generating multiple watermarks based on latent variable disturbance and noise scheduling diffusion. Background With the wide application of Diffusion Model in image generation, editing, cross-modal generation and other fields, the problems of authenticity and copyright protection of generated content are increasingly prominent. Traditional digital watermarking methods are mainly based on spatial or frequency domain embedding strategies, such as DCT, DWT or DCT-SVD, and the like, which have limited performance in resisting compression, regeneration or depth forgery attacks. In recent years, researchers have begun to explore watermark embedding using latent spatial properties of diffusion models, by encoding implicit information in latent variable or noise space, to achieve a more concealed, more robust watermarking scheme. However, the existing watermark research based on diffusion model mainly has the following problems: (1) The single watermark embedding limit capacity is that most methods only support single watermark information embedding, and multi-bit or multi-identity information is difficult to transmit simultaneously in the same generation process; (2) The traditional diffusion watermark is often embedded with information in the fixed noise intensity or diffusion step, the embedded intensity can not be dynamically adjusted according to the diffusion stage, and distortion or watermark failure is easily generated; (3) The existing method generally superimposes disturbance on the whole latent variable, but does not fully utilize semantic distribution characteristics of a diffusion model multi-channel latent space, so that interference exists among different information, and extraction accuracy is affected; (4) The robustness and reversibility are difficult to be combined, namely, the generation quality is destroyed by too strong embedding disturbance, and the effective watermark is difficult to be extracted after attack by too weak disturbance. Disclosure of Invention In order to solve the problems, the invention provides a diffusion generation multi-watermark method and a system based on latent variable disturbance and noise scheduling, which are used for realizing distributed parallel embedding of multi-bit watermarks by introducing a multi-channel latent variable watermark embedding module into a latent space of a diffusion model, and simultaneously, the self-adaptive noise scheduling module is combined to dynamically adjust the embedding amplitude and time position according to the diffusion step number and the signal to noise ratio, so that the robustness of the watermark is obviously enhanced while the high fidelity of an image is maintained. In order to achieve the above purpose, the invention provides a method and a system for generating multiple watermarks based on latent variable disturbance and noise scheduling diffusion. The method for generating the multiple watermarks based on the diffusion of the latent variable disturbance and the noise scheduling comprises the following steps: performing redundancy coding according to the input multi-bit watermark information to obtain a watermark coding sequence; According to the semantic characteristics of the watermark coding sequence and the latent variable channels, mapping the watermark coding sequence to a plurality of channels of diffusion model latent variables to obtain the latent variable channels with watermark codes; Dynamically adjusting noise scheduling parameters according to the number of diffusion steps and the signal-to-noise ratio to obtain a self-adaptive disturbance intensity control signal; Injecting disturbance with controllable amplitude into the latent variable channel with watermark coding according to the self-adaptive disturbance intensity control signal to obtain a latent variable with disturbance; Performing back diffusion sampling according to the latent variable with disturbance to obtain a generated image containing the watermark; performing latent variable inversion according to the generated image containing the watermark to obtain recovered latent variable characteristics; And detecting disturbance offset according to the restored latent variable characteristics, decoding, and extracting to obtain watermark information. Preferably, the process of obtaining the watermark coding sequence includes: Performing error correction coding or hash expansion according to the original watermark bit stream to obtain a coded sequence with error recovery capability; and performing sparse matrix mapping according to the coded sequence to obtain the watermark coding sequence. Preferably, the process of obtaining the watermarked latent variable channel