A hierarchical multi-domain cooperative migratable attack method for remote sensing target identification

By employing a hierarchical, multi-domain collaborative transferable adversarial attack method, and utilizing the collaborative perturbation of the frequency domain, spatial domain, and feature domain, the problem of poor transferability of adversarial samples in black-box attacks is solved, generating efficient and universal adversarial samples and improving the attack effect of remote sensing target recognition.

CN122156849APending Publication Date: 2026-06-05XIDIAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XIDIAN UNIV
Filing Date
2026-02-02
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing adversarial example generation methods are difficult to transfer effectively to black-box attack scenarios. Traditional methods suffer from overfitting, neglecting video domain and deep feature information, failing to decouple target semantics from background noise, and having large differences in the focus of different models on the target discrimination region, resulting in poor attack performance.

Method used

A hierarchical, multi-domain collaborative, transferable adversarial attack method is adopted. By combining frequency domain perturbation, spatial domain transformation, and feature domain perturbation with discrete wavelet transform, attention mechanism, and model-aware feature spatial perturbation, highly transferable adversarial examples are generated.

Benefits of technology

It significantly improves the transferability and attack success rate of adversarial examples across different model architectures, effectively decouples target semantics from background noise, and enhances the versatility and effectiveness of attacks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122156849A_ABST
    Figure CN122156849A_ABST
Patent Text Reader

Abstract

The application provides a hierarchical multi-domain collaborative migratable adversarial attack method for remote sensing target recognition, comprising: in each iteration, inputting an input image into a hierarchical multi-domain collaborative disturbance adversarial sample generation network to perform selective frequency domain disturbance on the input image, and then performing spatial transformation on the image after frequency domain disturbance to obtain an adversarial image; using a feature layer with added spatial disturbance to classify the adversarial image to obtain a classification result, and then updating the selective frequency domain disturbance and the disturbance added by the feature layer according to the loss between the classification result and the input image; repeating the process of generating the adversarial image until a constraint condition is reached to obtain a final adversarial sample. The application enriches the feature diversity of the adversarial sample from multiple dimensions by collaboratively integrating the semantic decoupling of the frequency domain, the modular transformation of the spatial domain and the model perception disturbance of the feature domain, thereby significantly improving the migration and attack success rate of the adversarial sample under the condition of black box attack for different architecture models.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Method for generating adversarial image

    CN108257116A

  • Black box adversarial sample generation method based on Transform model

    CN114693973A

  • Confrontation sample attack method based on adversarial disturbance transformation

    CN117011642A