计算机与数字工程2025，Vol.53Issue(2)：480-485,6.DOI:10.3969/j.issn.1672-9722.2025.02.032
基于条件生成对抗网络攻击目标检测模型的研究
A Study of Attack Target Detection Model Based on Conditional Generation Adversarial Network
摘要
Abstract
An attack model based on conditional generation of adversarial networks is designed with the aim of reducing the computational cost of the attack target detection model,increasing the attack strength,and reducing the perturbation.The model firstly incorporates a contrast learning approach based on an attention mechanism that assigns weights to pixels for capturing key fea-tures in the image,and contrast learning that shrinks the distance between key features of the adversarial sample and noisy features and expands the distance between key features of the clean sample.Secondly,this paper combines the joint attack method based on classifiers and regressors to mislead the classifier and regressor to make wrong judgments.The Faster-RCNN model is trained by us-ing PASCAL VOC 2007,VOC 2012 datasets.The experimental results demonstrate that the method reduces the mAP value of Fast-er-RCNN from 70.1%to 4.66%with an attack time of 0.04 s and an average perturbation of 5.39%under the VOC 2007 test set.Un-der the same test set,the adversarial sample migration attack SSD300 generated in Faster-RCNN,the mAP value decreases from 77.1%to 44.1%.关键词
目标检测/对抗攻击/生成对抗网络/注意力机制/对比学习Key words
target detection/adversarial attack/generating adversarial networks/attention mechanism/contrastive learning分类
计算机与自动化引用本文复制引用
周辉,琚贇,胡州明,祝文军..基于条件生成对抗网络攻击目标检测模型的研究[J].计算机与数字工程,2025,53(2):480-485,6.基金项目
国家重点研发计划项目(编号:2020YFB0905900)资助. （编号:2020YFB0905900）
