知识卡片
抽取相位影响输出
固定抽取网格只保留一部分位置,丢弃其余位置。具有相同视觉外观的高频微扰,在被保留位置上可能产生截然不同的滤波输出;书中玩具例子展示了这种敏感性。其结果受网格相位、滤波核、边界与扰动强度影响,不能直接推出所有模型都同样脆弱。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
The attacked input image is ℓin [n, m] + ϵz [n, m], where ϵ is the chosen adversarial strength. In this example we used ϵ = 0.05. As we can see in figure 21.1, the attacked input image is indistinguishable from the original input, ℓin [n, m], and yet their outputs are completely different. The culprit for this bizarre effect is the strided convolution that
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第21章《图像下采样与上采样》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c021.xhtml)
- 结论依据:摘录中关于抽取相位影响输出的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:The attacked input image is ℓin [n, m] + ϵz [n, m], where ϵ is the chosen adversarial strength. In this example we used ϵ = 0.05. As we can see in figure 21.1, the attacked input image is indistinguishable from the original input, ℓin [n, m], and yet their outputs are completely different. The culprit for this bizarre effect is the strided convolution that