Jailbreak Detection on LLaVA Vicuna-7B v1.6
92AccuracyKCD
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| KCDLayer=162025.12 | 92 | 94.1 | 10.1 | 92.2 | 97.7 | 97.2 | |
| MCDLayer=162025.12 | 91 | 97.2 | 15.2 | 91.6 | 98.6 | 98.8 | |
| KCDLayer=152025.12 | 89.4 | 95.3 | 16.5 | 90 | 96.9 | 96.3 | |
| KCDLayer=142025.12 | 89.1 | 94.9 | 16.6 | 89.8 | 96.4 | 96.2 | |
| MCDLayer=142025.12 | 88.3 | 95.5 | 18.9 | 89.1 | 97.4 | 97.5 | |
| MCDLayer=152025.12 | 88.3 | 96.9 | 20.3 | 89.2 | 98 | 98.1 | |
| JailDAM (RCS)Layer=CLIP2025.12 | 84.5 | 93.4 | 24.4 | 85.8 | 91.5 | 90 | |
| HiddenDetectLayer=16–292025.12 | 81.6 | 79.9 | 16.8 | 81.2 | 90.1 | 90 | |
| JailGuardLayer=Output2025.12 | 76.2 | 86 | 35.3 | 79.5 | 77.8 | 70.9 | |
| JailDAM (Original)Layer=CLIP2025.12 | 71.7 | 70.6 | 27.1 | 71.4 | 78.9 | 82.6 | |
| GradSafeLayer=Critical Param.2025.12 | 66.5 | 96.9 | 64.9 | 74.1 | 75.4 | 79.4 | |
| K-MeansLayer=FLAVA2025.12 | 62.3 | 97.4 | 72.9 | 72.1 | 59.8 | 62 | |
| LogisticLayer=FLAVA2025.12 | 59.8 | 37.6 | 18 | 48.3 | 61.7 | 66.8 |