知识卡片

普通读书笔记卡 · 9007.c · 生成模型:GPT-6

扰动球定义鲁棒范围

理解等级:高级

像素扰动鲁棒训练可写为 minf E(x,y)[max_(||ε||) L(f(x+ε),y)]。范数、半径、像素取值范围与标签保持假设共同定义威胁模型。对该集合改善的鲁棒性不能直接外推到换背景、语义变化或未知设备造成的全部分布漂移。

这张卡片看不懂?

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

      In the context of adversarial robustness, one form of equation (36.5) is to set T to be an epsilon-ball attack, that is, T(x) = x + ϵ: (36.6) This can increase robustness against the kinds of epsilon-ball attacks we saw in the previous chapter. Adversarial training has many other uses as well, beyond the scope of the current chapter: it is used for training generative

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml) - 结论依据:摘录中关于扰动球定义鲁棒范围的定义、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:In the context of adversarial robustness, one form of equation (36.5) is to set T to be an epsilon-ball attack, that is, T(x) = x + ϵ: (36.6) This can increase robustness against the kinds of epsilon-ball attacks we saw in the previous chapter. Adversarial training has many other uses as well, beyond the scope of the current chapter: it is used for training generative