Unverified50% confidenceFactExact time
Research indicates that an attacker needs to modify as little as 0.01% of training data to potentially implant backdoor behaviors in a model that are undetectable by standard evaluations.
1
Sources
50%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions74% similarUnverified成员推断攻击的核心假设是模型对训练集中的样本会产生更低的预测损失，通过设定损失阈值区分成员与非成员样本74% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.70% similarUnverified在标准行人检测数据集（如MOTChallenge）上训练的模型，在俯视视角下的检测精度（mAP）通常下降40%-60%69% similarUnverified卡内基梅隆大学研究显示对抗性图案攻击成功率在受控环境为90%以上，但引入正常视角变化后降至实际部署场景的30%以下69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/887229API
curl https://kongchang.com/api/v1/knowledge/claims/887229MCP
get_claim(id=887229)