Progressive Dual-Stage Enhancement for Universal Adversarial Patches
Jiale Xie, Ruo Wang, Shutai Wang, Teng Huang, Yuxi Zhao, Chenguang Xing
Chinese Aeronautical Establishment
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Existing techniques for Universal Adversarial Perturbations (UAP) have some limitations such as insufficient generalization, poor robustness in different domains and unstable optimization. To solve these problems, we have proposed the Progressive Dual-Stage Enhancement Framework (PDSE), which integrates an Adaptive Frequency Enhancer (AFE) for frequency-domain adaptive improvement and a Gradient-Guided Attention Enhancement module (GGAE). The AFE modifies the frequency distribution by emphasizing some specific frequencies, thus improving the essential low- and mid-frequency components for better adaptation between scenes while reducing the high-frequency noises. In contrast, the GGAE applies smoothed gradient saliency to establish joint channel-spatial attention, focusing on the parts which mainly affect the detection ability. Furthermore, a progressive activation scheme is introduced to improve the optimization stability and avoid premature overfitting. Experimental data from databases like INRIAPerson, CCTVPerson and COCOPerson show that PDSE significantly increases the attack efficiency and cross-domain robustness compared with the T-SEA baseline and conventional AdvPatch method. Additionally, PDSE performs well in generalizing to various detectors, including YOLOv3, YOLOv4 and YOLOv5.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIAdversarial Robustness in Machine Learning
Digital Media Forensic Detection · Advanced Neural Network Applications