Visual Question Answering on VQA v2 (Robust Accuracy)
90Robust AccuracyPDA
Evaluation Results
| Method | Links | |
|---|---|---|
| PDABase Model=GPT-4o, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 90 | |
| PDABase Model=Claude 3.5 Sonnet, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 88 | |
| PDABase Model=Gemini 1.5 Pro, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 86 | |
| Claude 3.5 SonnetDefense setting=No Defense, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 82 | |
| GPT-4oDefense setting=No Defense, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 81 | |
| PDAAttack=Transfer-based2026.04 | 81 | |
| PDAAttack=AnyAttack2026.04 | 81 | |
| PDAAttack=AttackVLM2026.04 | 79 | |
| No DefenseAttack=AnyAttack2026.04 | 79 | |
| No DefenseAttack=Transfer-based2026.04 | 77 | |
| Gemini 1.5 ProDefense setting=No Defense, Attack protocol=Transfer-attack (PGD on LLaVA)2026.04 | 76 | |
| No DefenseAttack=AttackVLM2026.04 | 75 |