Backdoor Detection on StrongREJECT prompts with triggers
100TPRWeightWatch
Evaluation Results
| Method | Links | |
|---|---|---|
| WeightWatchModel=trojan2, FPR threshold=< 2%2025.07 | 100 | |
| WeightWatchModel=trojan3, FPR threshold=< 2%2025.07 | 100 | |
| WeightWatchModel=trojan4, FPR threshold=< 2%2025.07 | 100 | |
| WeightWatchModel=trojan5, FPR threshold=< 2%2025.07 | 100 | |
| WeightWatchModel=trojan1, FPR threshold=< 2%2025.07 | 98.72 | |
| BEATModel=trojan3, FPR threshold=10%2025.07 | 97.76 | |
| Act. Diff.Model=trojan1, FPR threshold=2%2025.07 | 97.44 | |
| Act. Diff.Model=trojan2, FPR threshold=2%2025.07 | 96.49 | |
| BEATModel=trojan1, FPR threshold=10%2025.07 | 63.58 | |
| Act. Diff.Model=trojan4, FPR threshold=2%2025.07 | 51.76 | |
| BEATModel=trojan5, FPR threshold=10%2025.07 | 43.45 | |
| BEATModel=trojan3, FPR threshold=2%2025.07 | 31.95 | |
| BEATModel=trojan4, FPR threshold=10%2025.07 | 26.84 | |
| Act. Diff.Model=trojan5, FPR threshold=2%2025.07 | 19.81 | |
| Act. Diff.Model=trojan3, FPR threshold=2%2025.07 | 11.5 | |
| BEATModel=trojan5, FPR threshold=2%2025.07 | 7.99 | |
| BEATModel=trojan4, FPR threshold=2%2025.07 | 2.88 | |
| BEATModel=trojan1, FPR threshold=2%2025.07 | 1.92 | |
| BEATModel=trojan2, FPR threshold=10%2025.07 | 1.92 | |
| BEATModel=trojan2, FPR threshold=2%2025.07 | 0.32 |