A security brief this week describes a quiet risk in public datasets.
Security researchers say that if an attacker can tamper with that training data before the model is trained, the model may learn a hidden behavior.
The authors say the change can be hard to notice after training.
A security brief this week describes a quiet risk in public datasets.
1Security researchers say that
安全研究人员表示
2if an attacker can
如果攻击者能
4tamper with that training data
篡改那些训练数据
tamper with
暗中改动
3before the model is trained,
在模型训练之前
5the model may learn a hidden behavior.
模型可能会学到一种隐藏行为
安全研究人员表示,如果攻击者能在模型训练之前篡改那些训练数据,模型可能会学到一种隐藏行为。
The authors say the change can be hard to notice after training.