Targeted Attack on ViDoRe AI V1
100ASR-R@1 (mean)White-box (MO-PGD)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| White-box (MO-PGD)Embedder=CLIP-L, VLM=InternVL3-2B2025.04 | 100 | 100 | 100 | 100 | 21 | 0 | |
| White-box (MO-PGD)Embedder=CLIP-L, VLM=Qwen2.5-3B2025.04 | 100 | 100 | 89 | 100 | 22 | 0 | |
| White-box (MO-PGD)Embedder=CLIP-L, VLM=SmolVLM2025.04 | 100 | 100 | 98 | 100 | 23 | 0 | |
| In-set Model EnsembleEmbedder=CLIP-L, VLM=InternVL3-2B2025.04 | 100 | 100 | 53 | 65 | 21 | 0 | |
| In-set Model EnsembleEmbedder=CLIP-L, VLM=Qwen2.5-3B2025.04 | 100 | 100 | 81 | 100 | 22 | 1 | |
| In-set Model EnsembleEmbedder=CLIP-L, VLM=SmolVLM2025.04 | 100 | 100 | 57 | 74 | 23 | 1 | |
| Prompt-based (Gemini)Embedder=CLIP-L, VLM=InternVL3-2B2025.04 | 100 | 100 | 45 | 49 | 25 | — | |
| White-box (MO-PGD)Embedder=GME, VLM=InternVL3-2B2025.04 | 80 | 100 | 97 | 100 | 22 | 1 | |
| White-box (MO-PGD)Embedder=GME, VLM=SmolVLM2025.04 | 80 | 100 | 99 | 100 | 22 | 0 | |
| Component-wise TransferEmbedder=Same, VLM=Any2025.04 | 76 | 98 | 45 | 57 | 21 | 0 | |
| White-box (MO-PGD)Embedder=ColPali, VLM=InternVL3-2B2025.04 | 60 | 100 | 55 | 84 | 21 | 0 | |
| White-box (MO-PGD)Embedder=ColPali, VLM=SmolVLM2025.04 | 60 | 100 | 69 | 99 | 22 | 0 | |
| White-box (MO-PGD)Embedder=GME, VLM=Qwen2.5-3B2025.04 | 60 | 100 | 100 | 100 | 21 | 0 | |
| White-box (MO-PGD)Embedder=ColPali, VLM=Qwen2.5-3B2025.04 | 40 | 80 | 80 | 100 | 22 | 1 | |
| Prompt-based (GPT)Embedder=GME, VLM=SmolVLM2025.04 | 40 | 40 | 89 | 100 | 30 | — | |
| Out-set Model EnsembleEmbedder=Any, VLM=Any2025.04 | 0 | 0 | 47 | 56 | 21 | 0 | |
| Complete TransferEmbedder=Any, VLM=Any2025.04 | 0 | 0 | 46 | 56 | 21 | 0 | |
| Component-wise TransferEmbedder=Any, VLM=Same2025.04 | 0 | 0 | 85 | 100 | 21 | 0 |