List your company
ResearchFormerly WizWand

Backdoor Detection on StrongREJECT prompts with triggers

100TPRWeightWatch
-3.667223.246450.1677.0736Jul 31, 2025
Updated 6mo ago

Evaluation results

MethodLinks
WeightWatchModel=trojan2, FPR threshold=< 2%2025.07100
WeightWatchModel=trojan3, FPR threshold=< 2%2025.07100
WeightWatchModel=trojan4, FPR threshold=< 2%2025.07100
WeightWatchModel=trojan5, FPR threshold=< 2%2025.07100
WeightWatchModel=trojan1, FPR threshold=< 2%2025.0798.72
BEATModel=trojan3, FPR threshold=10%2025.0797.76
Act. Diff.Model=trojan1, FPR threshold=2%2025.0797.44
Act. Diff.Model=trojan2, FPR threshold=2%2025.0796.49
BEATModel=trojan1, FPR threshold=10%2025.0763.58
Act. Diff.Model=trojan4, FPR threshold=2%2025.0751.76
BEATModel=trojan5, FPR threshold=10%2025.0743.45
BEATModel=trojan3, FPR threshold=2%2025.0731.95
BEATModel=trojan4, FPR threshold=10%2025.0726.84
Act. Diff.Model=trojan5, FPR threshold=2%2025.0719.81
Act. Diff.Model=trojan3, FPR threshold=2%2025.0711.5
BEATModel=trojan5, FPR threshold=2%2025.077.99
BEATModel=trojan4, FPR threshold=2%2025.072.88
BEATModel=trojan1, FPR threshold=2%2025.071.92
BEATModel=trojan2, FPR threshold=10%2025.071.92
BEATModel=trojan2, FPR threshold=2%2025.070.32