PatchMap
收藏资源简介:
PatchMap 是第一个空间详尽的补丁放置基准,通过对 ImageNet 验证图像上的超过 10^8 次前向传递进行评估而构建。PatchMap 揭示了系统性的“热点”,在这些地方,小补丁(仅为图像的 2%)会诱导出自信的误分类并导致模型置信度大幅下降。PatchMap 旨在为研究人员提供一个空间详尽的基准,以便他们可以绘制漏洞热点图、训练位置感知检测器,并在相同条件下对优化或非优化的放置策略进行基准测试。
PatchMap is the first spatially exhaustive patch placement benchmark, developed by evaluating over 10^8 forward passes on ImageNet validation images. It reveals systematic "hotspots"—locations where small patches constituting merely 2% of an image’s total area induce confident misclassifications and lead to substantial reductions in model confidence. PatchMap aims to provide researchers with a spatially exhaustive benchmark to map vulnerability hotspots, train location-aware detectors, and benchmark both optimized and non-optimized placement strategies under identical experimental conditions.




