知识卡片
对抗漂移由对手选择
普通分布漂移来自环境变化;对抗漂移则允许对手选择输入以触发失败。威胁模型需要说明对方能修改什么、知道什么、能查询多少次及允许多大扰动。仅在自然测试集取得高准确率,不能推出这些条件下的安全性。
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 data shifts described previously may be disconcerting, but at least they are not malicious. Unfortunately, things can get worse. Quite often the data shift is not just random but adversarial. This may come up whenever our test setting is under the control of a competing agent, for example, someone trying to hack our system (computer security), another animal trying to eat us (evolution), or
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:摘录中关于对抗漂移由对手选择的定义、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 data shifts described previously may be disconcerting, but at least they are not malicious. Unfortunately, things can get worse. Quite often the data shift is not just random but adversarial. This may come up whenever our test setting is under the control of a competing agent, for example, someone trying to hack our system (computer security), another animal trying to eat us (evolution), or