An adversarial shape patch attack method for infrared target detection

By using Fourier series parameterization and rotation number integral to generate adversarial patches, the problem of insufficient shape representation capability in infrared target detection is solved, and stable missed detection and efficient attack of infrared detectors are achieved.

CN122244836APending Publication Date: 2026-06-19XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XI AN JIAOTONG UNIV
Filing Date
2026-04-15
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Among existing infrared target detection methods, those based on central ray star shape have insufficient shape representation capabilities, while those based on direct learning of masks from pixel grids have poor shape representation capabilities and fineness, resulting in poor attack performance.

Method used

The shape is parameterized by Fourier series, adversarial patch images are generated using the number of turns integral mapping, and high-frequency regularization is introduced to suppress noise. The Fourier coefficients are optimized to generate smooth and realistic adversarial patches.

Benefits of technology

It achieves stable false negatives against infrared detectors, improves shape characterization capabilities and precision, and the generated patches can effectively induce false negatives. Moreover, the attack effect is more stable under different confidence thresholds.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122244836A_ABST
    Figure CN122244836A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of artificial intelligence security technology and relates to an adversarial shape patch attack method for infrared target detection, including: 1. Parameter and model initialization: acquiring the original image, loading the infrared target detection model to be attacked, and locking the parameters; 2. Fourier shape parameter initialization and differentiable mapping of the number of turns to generate the initial patch; 3. Patch transformation and generation of adversarial image; 4. Inputting the adversarial image into the detector and obtaining the molecular set of target confidence scores for all candidate boxes; 5. Constructing a loss function and updating the Fourier coefficients using an optimizer; 6. Repeating the iteration a maximum number of times to obtain the optimized Fourier coefficients and obtain the final patch. This invention parameterizes the closed contour of the patch using Fourier series and combines the number of turns theorem to achieve differentiable mapping from the contour to the pixel mask, solving the problems of insufficient shape representation ability, poor precision, and dependence on explicit aggregation functions in the prior art.
Need to check novelty before this filing date? Find Prior Art