Visual Question Answering on VQA v2 ($L\_inf$ Perturbation Accuracy)
74.5Accuracy (Clean)No defense
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| No defenseVLM=LLaVA 1.5-7B2025.06 | 74.5 | 2.9 | 0 | |
| CLIPureVLM=LLaVA 1.5-7B2025.06 | 73.3 | 4.6 | 2.1 | |
| FARE2VLM=LLaVA 1.5-7B2025.06 | 71.7 | 34.9 | 23 | |
| DiffCAPVLM=LLaVA 1.5-7B2025.06 | 70.3 | 69.1 | 68.5 | |
| JPEG-DLVLM=LLaVA 1.5-7B2025.06 | 68.8 | 60.8 | 45.8 | |
| FARE4VLM=LLaVA 1.5-7B2025.06 | 68.3 | 40.7 | 30.5 | |
| DiffPureVLM=LLaVA 1.5-7B2025.06 | 67.3 | 65.8 | 66 | |
| TeCoA2VLM=LLaVA 1.5-7B2025.06 | 66.9 | 33.8 | 21.8 | |
| TeCoA4VLM=LLaVA 1.5-7B2025.06 | 63.2 | 41 | 31.7 | |
| No defenseVLM=OF-9B2025.06 | 48.5 | 1.8 | 0 | |
| CLIPureVLM=OF-9B2025.06 | 47.3 | 18.8 | 17.5 | |
| FARE2VLM=OF-9B2025.06 | 47 | 24 | 17.2 | |
| JPEG-DLVLM=OF-9B2025.06 | 46.8 | 39.5 | 32.4 | |
| DiffCAPVLM=OF-9B2025.06 | 46.3 | 45.4 | 45.3 | |
| TeCoA2VLM=OF-9B2025.06 | 46.2 | 23.5 | 20.5 | |
| FARE4VLM=OF-9B2025.06 | 46.1 | 23.6 | 21 | |
| DiffPureVLM=OF-9B2025.06 | 45.1 | 43.5 | 43.6 | |
| TeCoA4VLM=OF-9B2025.06 | 44.8 | 23.6 | 21.3 |