Censer.9 / reports /adversarial.md
KucLab's picture
Add report adversarial.md
bcfd5fd verified
|
Raw History Blame Contribute Delete
375 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 99.4% 0.917
typo 94.4% 0.858
leet 97.5% 0.906
homoglyph 100.0% 0.947
zero_width 100.0% 0.939
case 95.6% 0.833