An image detection adversarial sample generation method and system based on conditional diffusion

By combining frequency domain enhancement and semantic attention perturbation with an adversarial example generation method for image detection based on conditional diffusion, the shortcomings of existing methods in terms of transferability, image quality and robustness are addressed. The generated adversarial examples exhibit high transferability and stability in various detectors and are adapted to images generated by the diffusion model.

CN122116032APending Publication Date: 2026-05-29ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ZHEJIANG UNIV
Filing Date
2026-02-02
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Existing adversarial example generation methods suffer from poor transferability, image quality degradation, poor adaptability of images generated by diffusion models, and insufficient robustness when adversarial AIGI detection. They are particularly difficult to effectively bypass commercial detection systems in black-box scenarios.

Method used

We employ a conditional diffusion-based adversarial example generation method for image detection. By introducing small perturbations through integral gradients and combining frequency domain enhancement and semantic attention perturbations, we utilize a multi-dimensional guidance strategy to generate adversarial examples, including forward and reverse diffusion processes. We also use frequency exchange and diffusion constraint techniques to enhance the concealment and transferability of adversarial examples.

Benefits of technology

The generated adversarial examples exhibit high transferability and robustness across different environments and datasets, effectively bypassing multiple detectors, maintaining stable image quality, adapting to images generated by the diffusion model, and reducing frequency differences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122116032A_ABST
    Figure CN122116032A_ABST
Patent Text Reader

Abstract

The application discloses a kind of image detection based on condition diffusion Adversarial sample generation method and system, belong to artificial intelligence security field.Method includes: integral gradient method is used to original image is disturbed, after obtaining perturbed image, to perturbed image is gradually diffused step and is added noise, obtain final noise image and enter reverse diffusion, introduce conditionized correction noise estimation, based on current reverse diffusion step Adversarial intermediate sample, using DDIM, attention diagram extraction and frequency domain enhancement, respectively calculate to obtain loss L2, semantic dispersion loss and loss L4, further obtain total loss, further update Adversarial intermediate sample of current reverse diffusion step in reverse, and calculate next reverse diffusion step Adversarial intermediate sample according to updated Adversarial intermediate sample, until traverse all reverse diffusion steps, obtain final Adversarial sample.The Adversarial sample generated by the application not only improves the migration of adversarial perturbation, but also reduces the frequency difference of image.
Need to check novelty before this filing date? Find Prior Art