Visual Question Answering on TextVQA (Clean and Adversarial Accuracy)
0.371Accuracy (Clean)No defense
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| No defenseVLM=LLaVA 1.5-7B2025.06 | 0.371 | 0.005 | 0 | |
| CLIPureVLM=LLaVA 1.5-7B2025.06 | 0.361 | 0.021 | 0.014 | |
| JPEG-DLVLM=LLaVA 1.5-7B2025.06 | 0.346 | 0.272 | 0.211 | |
| FARE2VLM=LLaVA 1.5-7B2025.06 | 0.319 | 0.147 | 0.091 | |
| DiffCAPVLM=LLaVA 1.5-7B2025.06 | 0.283 | 0.29 | 0.289 | |
| FARE4VLM=LLaVA 1.5-7B2025.06 | 0.276 | 0.158 | 0.109 | |
| TeCoA2VLM=LLaVA 1.5-7B2025.06 | 0.241 | 0.121 | 0.088 | |
| No defenseVLM=OF-9B2025.06 | 0.238 | 0 | 0 | |
| JPEG-DLVLM=OF-9B2025.06 | 0.234 | 0.159 | 0.131 | |
| FARE2VLM=OF-9B2025.06 | 0.216 | 0.041 | 0.019 | |
| DiffPureVLM=LLaVA 1.5-7B2025.06 | 0.209 | 0.22 | 0.222 | |
| TeCoA4VLM=LLaVA 1.5-7B2025.06 | 0.207 | 0.126 | 0.093 | |
| CLIPureVLM=OF-9B2025.06 | 0.205 | 0.068 | 0.088 | |
| FARE4VLM=OF-9B2025.06 | 0.186 | 0.034 | 0.029 | |
| DiffCAPVLM=OF-9B2025.06 | 0.186 | 0.162 | 0.167 | |
| TeCoA2VLM=OF-9B2025.06 | 0.166 | 0.035 | 0.021 | |
| TeCoA4VLM=OF-9B2025.06 | 0.154 | 0.021 | 0.018 | |
| DiffPureVLM=OF-9B2025.06 | 0.136 | 0.132 | 0.135 |