PBGP Explorer
Perturbation-Based Generation Profiling — Interactive Results Dashboard
AI CONTROL HACKATHON 2026 · APART RESEARCH
6
Models
1,440
Generations
0.56–0.83
Features AUROC
0.94–1.00
PBGP AUROC
Method
FPR
4%
AUROC Gap — Features vs PBGP
Detection by Attack Type (MMD, 32B)
Operating Points