Censer.8 / reports /adversarial.md
KucLab's picture
Add report adversarial.md
7024ae0 verified
|
Raw History Blame Contribute Delete
376 Bytes

Adversarial robustness (unsafe items under text attacks)

n=160 unsafe items, threshold 0.5, detection rate = fraction still flagged unsafe.

Attack Detection rate Mean unsafe prob
clean 100.0% 0.890
typo 92.5% 0.826
leet 96.9% 0.869
homoglyph 100.0% 0.922
zero_width 100.0% 0.914
case 96.2% 0.817