Adversarial sample generation method based on multi-level salient region knowledge distillation

The adversarial example generation method based on multi-level salient region knowledge distillation addresses the shortcomings of existing full-image adversarial attacks, improves the success rate and concealment of attacks, and enhances the generation efficiency and generalization ability of adversarial examples.

CN121146015BActive Publication Date: 2026-06-09CHINESE PEOPLES LIBERATION ARMY UNIT 32801
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
CHINESE PEOPLES LIBERATION ARMY UNIT 32801
Filing Date
2025-09-05
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing full-image adversarial attacks suffer from large perturbation amplitudes that are easily detected, low attack efficiency due to ignoring multi-scale semantic information, and insufficient consideration of the significant differences between different levels of the model, making it difficult to balance attack success rate and stealth.

Method used

The adversarial example generation method based on multi-level salient region knowledge distillation generates a multi-level salient region image set by locating key pixels of the neural network, and integrates the region images using knowledge distillation of the salient image set to generate a high-quality adversarial attack.

Benefits of technology

It improves the efficiency of adversarial sample generation and attack success rate, enhances the stealth and generalization ability of attacks, and comprehensively considers the significant differences at different levels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121146015B_ABST
    Figure CN121146015B_ABST
Patent Text Reader

Abstract

The present application relates to a kind of based on multi-level saliency region knowledge distillation's adversarial sample generation method, belong to the field of computer vision of adversarial attack technique field.The present application utilizes multi-level saliency region information guide, locates the key pixel that neural network makes judgment to single image, obtains the saliency region image of each feature layer, constitutes multi-level saliency image set;Again, using the strategy of saliency image set knowledge distillation, integrates the saliency region image in saliency image set, to generate higher quality of adversarial attack.The present application comprehensively considers the saliency difference of different levels, improves attack success rate and concealment;The present application designs the saliency region integration strategy based on data set distillation, improves the generalization ability of adversarial attack.
Need to check novelty before this filing date? Find Prior Art