Image Captioning on MS-COCO 2017 (val)
55.4ASR (LLaVA*)V-Attack
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| V-AttackTrain=Single-surrogate, Data Augmentation=Yes2025.11 | 55.4 | 28.3 | 21 | 44.5 | 35.4 | 3 | 3.8 | |
| V-AttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 50.4 | 53.6 | 56 | 66.8 | 56.7 | 7.4 | 7.9 | |
| M-AttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 37 | 40.5 | 48.3 | 54.4 | 45 | 7.9 | 8.5 | |
| SSA-CWATrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 26.2 | 30.4 | 24.1 | 28.5 | 27.3 | 8.5 | 9.1 | |
| MF-iiTrain=Single-surrogate, Data Augmentation=No2025.11 | 10.3 | 12.5 | 7.3 | 3.9 | 11.4 | 3.2 | 3.9 | |
| MF-itTrain=Single-surrogate, Data Augmentation=No2025.11 | 5.1 | 4 | 4 | 2.8 | 4.6 | 3.6 | 4.1 | |
| AdvDiffTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 4.3 | 3.7 | 3.7 | 3.1 | 3.9 | 10.6 | 11.3 | |
| AnyAttackTrain=Ensemble-surrogate, Data Augmentation=Yes2025.11 | 4.2 | 5 | 4.3 | 3 | 4.5 | 8.8 | 10.1 |