Visual Question Answering on TextVQA (VQA accuracy)
39VQA AccuracyClean
Evaluation Results
| Method | Links | |
|---|---|---|
| CleanModel=LLaVa1.5-13B2026.02 | 39 | |
| CleanModel=LLaVa1.5-7B2026.02 | 37.1 | |
| VT-AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.02 | 28.7 | |
| VT-AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.02 | 26.9 | |
| MIX.AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.02 | 24.6 | |
| CleanModel=OF-9B2026.02 | 23.8 | |
| MIX.AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.02 | 22.8 | |
| AttackVLM-iiModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.02 | 19.7 | |
| MIX.AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.02 | 19.7 | |
| AttackVLM-iiModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.02 | 19.6 | |
| MIX.AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.02 | 19.1 | |
| VT-AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.02 | 18.4 | |
| VT-AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.02 | 16.1 | |
| VT-AttackModel=OF-9B, Perturbation (ε)=2/2552026.02 | 16.1 | |
| AttackVLM-iiModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.02 | 14.7 | |
| AttackVLM-iiModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.02 | 14 | |
| VEAttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.02 | 13.8 | |
| MIX.AttackModel=OF-9B, Perturbation (ε)=2/2552026.02 | 13.4 | |
| VT-AttackModel=OF-9B, Perturbation (ε)=4/2552026.02 | 13.2 | |
| VEAttackModel=OF-9B, Perturbation (ε)=2/2552026.02 | 12.5 | |
| VEAttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.02 | 12.4 | |
| VEAttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.02 | 10.1 | |
| AttackVLM-iiModel=OF-9B, Perturbation (ε)=2/2552026.02 | 10 | |
| MIX.AttackModel=OF-9B, Perturbation (ε)=4/2552026.02 | 8.8 | |
| VEAttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.02 | 8.6 | |
| PA-AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.02 | 8.3 | |
| AttackVLM-iiModel=OF-9B, Perturbation (ε)=4/2552026.02 | 7.4 | |
| PA-AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.02 | 7 | |
| PA-AttackModel=OF-9B, Perturbation (ε)=2/2552026.02 | 6.6 | |
| PA-AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.02 | 6.5 | |
| VEAttackModel=OF-9B, Perturbation (ε)=4/2552026.02 | 5.7 | |
| PA-AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.02 | 5.1 | |
| PA-AttackModel=OF-9B, Perturbation (ε)=4/2552026.02 | 4.8 |