知识卡片
成员推断与训练提取目标不同
成员推断判断某段数据是否参与训练,训练数据提取则尝试恢复训练内容。二者对应不同隐私威胁与评估目标;黑盒可访问不表示一定成功或可对每条数据做确定判断,结果需考虑误报和审计范围。
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
- · 预训练数据
原书出处 · 1 处
- · 查看原书依据
Membership inference attack With just closed-box access to a model, a membership inference attack enables an attacker to determine if a sequence of text has been used to train the model or not. Training data extraction attack With just closed-box access to a model, the attacker can prompt the model to generate memorized sensitive information. A naive example involves prompting the model with the text
参考来源
- 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html)
- 结论依据:摘录中关于成员推断与训练提取目标不同的定义、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:Membership inference attack With just closed-box access to a model, a membership inference attack enables an attacker to determine if a sequence of text has been used to train the model or not. Training data extraction attack With just closed-box access to a model, the attacker can prompt the model to generate memorized sensitive information. A naive example involves prompting the model with the text