Adaptive multi-modal adversarial code image generation method and device
By adopting an adaptive multimodal adversarial CAPTCHA image generation method, combining text and image-level adversarial generation and utilizing environmental risk perception to adjust the perturbation intensity, the problem of existing CAPTCHA images being easily decrypted is solved, thus improving the defense capability of adversarial CAPTCHAs.
CN122289447APending Publication Date: 2026-06-26ANHUI UNIV OF SCI & TECH
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV OF SCI & TECH
- Filing Date
- 2026-04-14
- Publication Date
- 2026-06-26
AI Technical Summary
Technical Problem
Existing adversarial CAPTCHA images are easily dissolved by diffusion models, which weakens the defense capability against multimodal large models.
Method used
An adaptive multimodal adversarial CAPTCHA image generation method is adopted, which combines implicit adversarial instructions at the text level with adversarial guidance at the image level, and introduces intensity adjustment of environmental risk perception to generate CAPTCHA images.
Benefits of technology
It significantly improves the defense capability against multimodal large-scale CAPTCHA image cracking behavior, avoids the removal of perturbations in the reverse denoising process of diffusion model, and enhances robustness and concealment.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN122289447A_ABST
Abstract
This disclosure provides an adaptive multimodal adversarial CAPTCHA image generation method and apparatus, comprising: obtaining the risk level of the current CAPTCHA image usage environment; determining the text perturbation intensity and image perturbation intensity required for CAPTCHA image generation based on the risk level; embedding implicit adversarial instructions into the natural language instruction text used for CAPTCHA image generation based on the text perturbation intensity to induce logical contradictions or illusions in the target model; the target model being a multimodal large model used to crack the CAPTCHA image; and inputting the natural language instruction text with embedded implicit adversarial instructions and the image perturbation intensity into a pre-trained latent space diffusion model, which then generates the CAPTCHA image. In this way, this disclosure effectively solves the problem that existing adversarial CAPTCHA images are easily neutralized by diffusion models, significantly improving the defense capability of adversarial CAPTCHA images against multimodal large model cracking behavior.
Need to check novelty before this filing date? Find Prior Art