Search

CN-122020619-A - Multi-mode generation type artificial intelligent watermark embedding and tracing method

CN122020619ACN 122020619 ACN122020619 ACN 122020619ACN-122020619-A

Abstract

The invention discloses a multi-mode generation type artificial intelligent watermark embedding and tracing method, which comprises the steps of constructing a three-order nonlinear hybrid power system with fractional order Caputo derivative and time-varying Rickettsia matrix closed loop feedback, constructing a regular Vine structure which at least comprises three layers of R-Vine, C-Vine and D-Vine, wherein the total number of tree nodes is not less than 4096, generating an opposite watermark disturbance field through sphere radial-Vine Copula inverse mapping by a finite time chaotic track of the hybrid power system, and simultaneously injecting the disturbance field into an ultrahigh-dimensional singular value manifold of a generation type artificial intelligent output content according to a nonlinear variation embedding mode. The method of the invention generates a strong anti-interference watermark disturbance field through the combination of the chaotic system and the rattan structure, realizes the effective watermark embedding and reliable tracing of the multi-mode AI content, and ensures the attribution of intellectual property rights.

Inventors

  • ZHAO TIANCHENG
  • WANG HAO
  • YU HAI
  • WANG SHIWEI
  • LI CHENG

Assignees

  • 杭州联汇科技股份有限公司

Dates

Publication Date
20260512
Application Date
20260107

Claims (6)

  1. 1. A multi-mode generation type artificial intelligent watermark embedding and tracing method is characterized by comprising the following steps: Step one, constructing a third-order nonlinear hybrid power system with a fractional order Caputo derivative and a time-varying Rickettsia matrix closed-loop feedback; step two, constructing a regular Vine structure at least comprising three layers of nesting of R-Vine, C-Vine and D-Vine, wherein the total number of tree nodes is not less than 4096; and thirdly, generating an antagonistic watermark disturbance field through the limited-time chaotic track of the hybrid power system by means of sphere radial-vine Copula inverse mapping, and simultaneously injecting the disturbance field into an ultra-high-dimensional singular value manifold of the generated artificial intelligent output content in a nonlinear variation embedding mode.
  2. 2. The method for embedding and tracing a multi-modal generated artificial intelligent watermark as set forth in claim 1, wherein the time domain dynamic trace of the fractional order lyapunov-licarpa hybrid system watermark in step one is generated by the following fractional order nonlinear coupling differential equation: Wherein: The binary watermark bit is chaos excitation after BCH-vine Copula encoding, wherein the binary watermark bit is the fraction derivative of Caputo, the order alpha epsilon (0.9,1), beta epsilon (0.8,0.9), gamma epsilon (0.95,1), P (t) is a time-varying Richka matrix, Q is a positive weighting matrix, and u (t) is the chaos excitation after binary watermark bit is encoded by the BCH-vine Copula.
  3. 3. The method for embedding and tracing a multi-modal generated artificial intelligent watermark as claimed in claim 2, wherein the system in the first step presents a finite time chaotic attractor under specific parameters, and the maximum Lyapunov exponent satisfies: 。
  4. 4. The method for embedding and tracing a multi-modal generated artificial intelligent watermark as set forth in claim 3, wherein said ultra-high-dimensional rattan Copula structure in step two uses the trajectory vector generated by said power system Mapping to rattan Copula space: Wherein, each layer of pair-copula selects 27 parameter transformation BB8 group + Tawntype-2 asymmetric group mixed core, and the parameters are estimated layer by layer on 4096-dimensional historical tracks through maximum likelihood.
  5. 5. The method for embedding and tracing a multi-modal generated artificial intelligent watermark according to claim 4, wherein the antagonistic embedding transformation in the second step is to perform spherical radial integration on a rattan Copula density function, and obtain an embedding strength field epsilon (x, y) through a variable component sub-modal invisible state transfer simulator (VQST-simulator), and the final watermark embedding formula is as follows: The perturbation has a PSNR >48dB on the human visual system, but can carry 4096bit trace information on a 4096×4096 large map.
  6. 6. The method of claim 5, further comprising the steps of extracting and tracing, wherein the extracting end extracts by solving a 1024-dimensional non-convex Likati-vine Copula inverse problem, and the theoretical bit error rate is < 10-27.

Description

Multi-mode generation type artificial intelligent watermark embedding and tracing method Technical Field The invention relates to the technical field of multi-mode generation type artificial intelligent watermarks, in particular to a multi-mode generation type artificial intelligent watermark embedding and tracing method. Background With the wide application of the generated Artificial Intelligence (AIGC) content, the problems of difficult tracing and easy malicious tampering of multi-mode content such as images, videos, audios and texts are increasingly highlighted, and the method brings serious challenges to the copyright protection, information security and responsibility identification of the content. The traditional digital watermarking technology mostly adopts methods such as low-dimensional Gaussian Copula or simple frequency domain transformation, and the like, has obvious limitations in coping with deep counterfeiting attacks and high-intensity compression, and is difficult to effectively consider both robustness and imperceptibility of the watermark due to poor distribution matching property on a high-dimensional manifold, so that the comprehensive requirements of the current AIGC content governance on traceability, anti-attack performance and user experience cannot be met. In addition, the prior art lacks a unified standard framework support, further restricts the large-scale application of the system in an actual scene, and is difficult to realize effective identification and responsibility locking of AIGC contents. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide a multi-mode generation type artificial intelligent watermark embedding and tracing method, which is used for generating an antagonistic watermark disturbance field and embedding the antagonistic watermark disturbance field into a high-dimensional manifold of AIGC content by constructing a fractional order chaotic power system and an ultra-high-dimensional rattan Copula structure, so as to realize the compatibility of robustness and imperceptibility and solve the problems of weak anti-attack capability and poor distribution matching of the traditional watermark technology. In order to achieve the purpose, the invention provides a multi-mode generation type artificial intelligent watermark embedding and tracing method, which comprises the following steps: Step one, constructing a third-order nonlinear hybrid power system with a fractional order Caputo derivative and a time-varying Rickettsia matrix closed-loop feedback; step two, constructing a regular Vine structure at least comprising three layers of nesting of R-Vine, C-Vine and D-Vine, wherein the total number of tree nodes is not less than 4096; and thirdly, generating an antagonistic watermark disturbance field through the limited-time chaotic track of the hybrid power system by means of sphere radial-vine Copula inverse mapping, and simultaneously injecting the disturbance field into an ultra-high-dimensional singular value manifold of the generated artificial intelligent output content in a nonlinear variation embedding mode. As a further improvement of the invention, the time domain dynamic track of the watermark of the fractional order Lyapunov-Rickettsia hybrid power system in the step one is generated by the following fractional order nonlinear coupling differential equation: Wherein: The binary watermark bit is chaos excitation after BCH-vine Copula encoding, wherein the binary watermark bit is the fraction derivative of Caputo, the order alpha epsilon (0.9,1), beta epsilon (0.8,0.9), gamma epsilon (0.95,1), P (t) is a time-varying Richka matrix, Q is a positive weighting matrix, and u (t) is the chaos excitation after binary watermark bit is encoded by the BCH-vine Copula. As a further improvement of the invention, the system in the first step presents a finite time chaotic attractor under specific parameters, and the maximum Lyapunov exponent of the attractor meets the following conditions: 。 As a further improvement of the invention, the ultra-high-dimensional rattan Copula structure in the second step is used for carrying out the track vector generated by the power system Mapping to rattan Copula space: Wherein, each layer of pair-copula selects 27 parameter transformation BB8 group + Tawntype-2 asymmetric group mixed core, and the parameters are estimated layer by layer on 4096-dimensional historical tracks through maximum likelihood. The invention is further improved in that after the antagonistic embedding transformation in the second step carries out spherical radial integration on the vine Copula density function, an embedding strength field epsilon (x, y) is obtained through a variable component sub-state invisible state transmission simulator (VQST-simulator), and a final watermark embedding formula is as follows: The perturbation has a PSNR >48dB on the human visual system, but can carry 4096bit trace inform