Jailbreak Defense on Qwen2-VL
0ASRBenign image
Evaluation Results
| Method | Links | |
|---|---|---|
| Benign imageepsilon=16/2552026.03 | 0 | |
| Benign imageepsilon=32/2552026.03 | 0 | |
| Benign imageepsilon=64/2552026.03 | 0 | |
| Benign imageepsilon=unconstrained2026.03 | 0 | |
| NullSteerepsilon=16/2552026.03 | 4.21 | |
| NullSteerepsilon=32/2552026.03 | 4.55 | |
| ASTRAepsilon=32/2552026.03 | 5 | |
| ASTRAepsilon=16/2552026.03 | 6.06 | |
| NullSteerepsilon=unconstrained2026.03 | 13.25 | |
| JailGuardepsilon=unconstrained2026.03 | 15.45 | |
| ASTRAepsilon=unconstrained2026.03 | 15.45 | |
| NullSteerepsilon=64/2552026.03 | 15.82 | |
| ASTRAepsilon=64/2552026.03 | 18.18 | |
| JailGuardepsilon=16/2552026.03 | 19.09 | |
| JailGuardepsilon=32/2552026.03 | 20 | |
| JailGuardepsilon=64/2552026.03 | 21.82 | |
| Refusal Pairsepsilon=64/2552026.03 | 21.82 | |
| ECSOepsilon=32/2552026.03 | 27.27 | |
| Refusal Pairsepsilon=16/2552026.03 | 29.09 | |
| ECSOepsilon=16/2552026.03 | 30 | |
| ECSOepsilon=64/2552026.03 | 31.82 | |
| Refusal Pairsepsilon=32/2552026.03 | 31.82 | |
| ECSOepsilon=unconstrained2026.03 | 32.73 | |
| Self-reminderepsilon=64/2552026.03 | 40 | |
| Self-reminderepsilon=32/2552026.03 | 47.27 | |
| Self-reminderepsilon=16/2552026.03 | 50 | |
| Refusal Pairsepsilon=unconstrained2026.03 | 52.73 | |
| Self-reminderepsilon=unconstrained2026.03 | 58.18 | |
| Jailbreak Templatesepsilon=64/2552026.03 | 65.45 | |
| w/o defenseepsilon=16/2552026.03 | 67.27 | |
| Jailbreak Templatesepsilon=16/2552026.03 | 68.18 | |
| Jailbreak Templatesepsilon=32/2552026.03 | 68.18 | |
| w/o defenseepsilon=32/2552026.03 | 70.46 | |
| w/o defenseepsilon=64/2552026.03 | 71.82 | |
| Jailbreak Templatesepsilon=unconstrained2026.03 | 74.55 | |
| w/o defenseepsilon=unconstrained2026.03 | 76.36 |