Skip to content

The monitor as a noisy classifier: ROC & AUC

Model the trusted monitor's suspicion scores as two overlapping distributions. Drag its quality to see the ROC curve, AUC, and how many backdoors slip past a fixed 2% false-positive threshold.

Innocent solutions Backdoored solutions

Suspicion-score distributions

audit →more suspicious →

ROC curve

AUC 0.8770FPR 1TPR

AUC

0.877

Backdoors caught @ 2% FPR

34.0%

Missed (false negatives)

66.0%