Visual Question Answering on MS-COCO 2017 (val)
54.2ASR (LLaVA*)V-Attack
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| V-AttackTrain=Single-surrogate, Data Augmentation=Yes2025.11 | 54.2 | 35.2 | 24 | 39.1 | 35.5 | 0.03 | 0.038 | |
| V-AttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 45.3 | 55.5 | 63.6 | 59.7 | 56 | 0.074 | 0.079 | |
| M-AttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 36.3 | 40.9 | 39.8 | 47.5 | 41.1 | 0.079 | 0.085 | |
| SSA-CWATrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 22.9 | 31.2 | 23.3 | 27 | 27 | 0.085 | 0.091 | |
| MF-iiTrain=Single-surrogate, Data Augmentation=No2025.11 | 6.6 | 13 | 6 | 4.5 | 7.8 | 0.032 | 0.039 | |
| AnyAttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 3.7 | 10 | 4.5 | 3 | 5.8 | 0.088 | 0.101 | |
| AdvDiffTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 2.9 | 6.1 | 6 | 2.8 | 5.4 | 0.106 | 0.113 | |
| MF-itTrain=Single-surrogate, Data Augmentation=No2025.11 | 2.6 | 3.5 | 3.1 | 3.3 | 3.1 | 0.036 | 0.041 |