Visual Question Answering on OKVQA (VQAScore, ASR)
73.07ASR (Success Rate)Phantasia
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PhantasiaInputs=Poisoned2026.04 | 73.07 | — | |
| BadVLMInputs=Poisoned2026.04 | 72.66 | — | |
| VLOODInputs=Poisoned2026.04 | 70.09 | — | |
| TrojVLMInputs=Poisoned2026.04 | 69.41 | — | |
| PhantasiaModel=BLIP2, Inputs=Poisoned2026.04 | 68.5 | — | |
| ShadowcastInputs=Poisoned2026.04 | 68.17 | — | |
| BadVLMModel=BLIP2, Inputs=Poisoned2026.04 | 65.51 | — | |
| ShadowcastModel=BLIP2, Inputs=Poisoned2026.04 | 63.12 | — | |
| VLOODModel=BLIP2, Inputs=Poisoned2026.04 | 63.1 | — | |
| TrojVLMModel=BLIP2, Inputs=Poisoned2026.04 | 62.91 | — | |
| PhantasiaModel=LLaVA, Inputs=Poisoned2026.04 | 43.12 | — | |
| VLOODModel=LLaVA, Inputs=Poisoned2026.04 | 42.52 | — | |
| ShadowcastModel=LLaVA, Inputs=Poisoned2026.04 | 42.26 | — | |
| TrojVLMModel=LLaVA, Inputs=Poisoned2026.04 | 42.16 | — | |
| BadVLMModel=LLaVA, Inputs=Poisoned2026.04 | 41.05 | — | |
| TrojVLMModel=BLIP2, Inputs=Clean2026.04 | 3.62 | 27.52 | |
| ShadowcastModel=BLIP2, Inputs=Clean2026.04 | 3.55 | 27.14 | |
| VLOODModel=BLIP2, Inputs=Clean2026.04 | 3.52 | 28.31 | |
| PhantasiaModel=BLIP2, Inputs=Clean2026.04 | 3.5 | 39.55 | |
| TrojVLMInputs=Clean2026.04 | 2.61 | 31.63 | |
| BadVLMModel=BLIP2, Inputs=Clean2026.04 | 2.54 | 32.88 | |
| VLOODInputs=Clean2026.04 | 2.4 | 32.42 | |
| ShadowcastInputs=Clean2026.04 | 2.05 | 30.76 | |
| PhantasiaInputs=Clean2026.04 | 1.91 | 34.45 | |
| BadVLMInputs=Clean2026.04 | 1.5 | 33.62 | |
| TrojVLMModel=LLaVA, Inputs=Clean2026.04 | 1.32 | 34.44 | |
| PhantasiaModel=LLaVA, Inputs=Clean2026.04 | 1.31 | 35.52 | |
| BadVLMModel=LLaVA, Inputs=Clean2026.04 | 1.28 | 34.23 | |
| ShadowcastModel=LLaVA, Inputs=Clean2026.04 | 1.27 | 34.55 | |
| VLOODModel=LLaVA, Inputs=Clean2026.04 | 1.13 | 35.16 |