知识卡片

普通读书笔记卡 · 9004.b · 生成模型:GPT-6

定向扰动应最小化目标损失

理解等级:高级

以目标类别 t 的交叉熵定义 H,定向攻击可写为 min_ε H(y_t,f(x+ε)),约束 ||ε||;无定向攻击则可增大真实类别损失。书中式 35.6 印为对目标类别交叉熵取 arg max,与其“猫变鸵鸟”意图相反,应纠正方向。小范数也不自动保证视觉不可察觉,成功率依赖模型与约束。

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

      This was famously demonstrated by [461], who showed that one can add imperceptible noise to an image that fools neural net classifiers despite looking completely innocuous to a human. In fact, the noise pattern can be selected to force the network to produce any output the attacker desires. These are called adversarial attacks because we suppose the attacker, who adds the noise, is an adversary

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(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/c035b.xhtml) - 结论依据:摘录中关于定向扰动应最小化目标损失的定义、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:This was famously demonstrated by [461], who showed that one can add imperceptible noise to an image that fools neural net classifiers despite looking completely innocuous to a human. In fact, the noise pattern can be selected to force the network to produce any output the attacker desires. These are called adversarial attacks because we suppose the attacker, who adds the noise, is an adversary