Unverified50% confidenceFactExact time
数据投毒攻击按攻击目标可分为可用性攻击和目标攻击，前者旨在降低模型整体性能，后者只在特定触发条件下改变模型行为
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
抗投毒概念锚定：防御AI数据污染的新思路
Related Claims
VerifiedData poisoning attacks involve mixing malicious samples into training datasets, causing the model to exhibit incorrect behavior toward specific inputs after deployment.72% similarUnverified工具调用使攻击面从让模型说错话升级为让模型做错事，危害从信息误导上升为实质性的数据访问和系统篡改71% similarUnverified下游用户对模型的二次微调可能意外破坏原有安全特性67% similarUnverified逆向分析中通过全局搜索区分服务端返回值与前端生成值，是攻击面最小化原则的应用66% similarUnverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/772959API
curl https://kongchang.com/api/v1/knowledge/claims/772959MCP
get_claim(id=772959)