← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6618.c · 生成模型:GPT-6

隐私措施需区分风险降低与保证

理解等级:高级

清洗、解码限制与事后过滤可降低某些泄漏风险,但不能自动提供形式化隐私保证。差分隐私等保证依赖其具体机制、参数与会计条件;不能把任意加噪或一次攻击未成功称为已获保证,性能取舍也需实测。

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

      only one of several solutions to prevent data leakage from models. One promising technique is differential privacy, which introduces randomness in the inputs or outputs to provide theoretical guarantees for privacy preservation. In neural networks, differential privacy is implemented using the DP-SGD algorithm, which involves gradient clipping and noise addition at the end of each update. However, differential privacy significantly slows training, negatively affects model

    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:摘录中关于隐私措施需区分风险降低与保证的定义、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:only one of several solutions to prevent data leakage from models. One promising technique is differential privacy, which introduces randomness in the inputs or outputs to provide theoretical guarantees for privacy preservation. In neural networks, differential privacy is implemented using the DP-SGD algorithm, which involves gradient clipping and noise addition at the end of each update. However, differential privacy significantly slows training, negatively affects model