Image Captioning Robustness on Image Captioning Dataset
82.9CLIP Score (RN-50)Chain of Attack
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Chain of AttackTarget VLM=VIECAP [19]2024.11 | 82.9 | 81.9 | 83.8 | 84.7 | 78.2 | 82.3 | 98.4 | 99.5 | |
| MF-itTarget VLM=VIECAP [19]2024.11 | 78 | 76.7 | 78.9 | 79.6 | 71.8 | 77 | 69.3 | 79.9 | |
| MF-iiTarget VLM=VIECAP [19]2024.11 | 76.4 | 75.3 | 77.4 | 78 | 70.1 | 75.4 | 76.6 | 85.8 | |
| Chain of AttackTarget VLM=UNIDIFFUSER [4]2024.11 | 76.1 | 74.4 | 77.2 | 78.5 | 69.8 | 75.2 | 94.2 | 98.9 | |
| MF-iiTarget VLM=UNIDIFFUSER [4]2024.11 | 70.9 | 69.5 | 72.1 | 73.3 | 63.7 | 70 | 90 | 98.4 | |
| Chain of AttackTarget VLM=SMALLCAP [42]2024.11 | 68.6 | 66.1 | 70 | 71.1 | 60.4 | 67.2 | 56.8 | 66.5 | |
| MF-iiTarget VLM=SMALLCAP [42]2024.11 | 67.3 | 65 | 68.5 | 69.8 | 58.6 | 65.8 | 47.1 | 52.2 | |
| MF-itTarget VLM=UNIDIFFUSER [4]2024.11 | 65.5 | 63.9 | 67.8 | 69.8 | 61.1 | 65.6 | 80.2 | 95.8 | |
| MF-itTarget VLM=SMALLCAP [42]2024.11 | 57.9 | 54.8 | 59.1 | 60.7 | 46.6 | 55.8 | 22.1 | 26.8 | |
| AttackBardTarget VLM=SMALLCAP [42]2024.11 | 53.2 | 48.4 | 51.5 | 56.6 | 39.2 | 49.8 | 6.6 | 9.2 | |
| Mix.AttackTarget VLM=SMALLCAP [42]2024.11 | 52.9 | 48.3 | 51.5 | 56.4 | 39.2 | 49.7 | 5.8 | 8.1 | |
| AttackBardTarget VLM=UNIDIFFUSER [4]2024.11 | 52.2 | 48.6 | 53.1 | 56.5 | 56.5 | 53.4 | 8.1 | 14.4 | |
| Chain of AttackTarget VLM=LLAVA-7B [29]2024.11 | 51.1 | 49.6 | 52 | 55.2 | 35.8 | 48.7 | 14.5 | 28.4 | |
| Clean imageTarget VLM=SMALLCAP [42]2024.11 | 50.7 | 48.6 | 51.1 | 52.7 | 37.5 | 48.1 | — | — | |
| Mix.AttackTarget VLM=VIECAP [19]2024.11 | 49.6 | 47 | 48.8 | 52.1 | 36.7 | 46.8 | 11.1 | 17.4 | |
| AttackBardTarget VLM=VIECAP [19]2024.11 | 49.2 | 46.7 | 48.7 | 51.7 | 36.3 | 46.5 | 15.5 | 25 | |
| Chain of AttackTarget VLM=LLAVA-13B [29]2024.11 | 48.1 | 48 | 49.4 | 49 | 34.6 | 45.8 | 12.3 | 24.3 | |
| AttackBardTarget VLM=LLAVA-7B [29]2024.11 | 47.9 | 47.4 | 48.1 | 48.5 | 34.6 | 45.3 | 2 | 3.7 | |
| AttackBardTarget VLM=LLAVA-13B [29]2024.11 | 47.9 | 47.4 | 48.1 | 48.5 | 34.6 | 45.3 | 2.6 | 4.8 | |
| MF-iiTarget VLM=LLAVA-13B [29]2024.11 | 47.4 | 47.2 | 48.7 | 48.4 | 34.4 | 45.2 | 3.6 | 6.9 | |
| MF-iiTarget VLM=LLAVA-7B [29]2024.11 | 47.2 | 46.7 | 48.2 | 48 | 34.2 | 44.9 | 2.6 | 4.7 | |
| Clean imageTarget VLM=LLAVA-7B [29]2024.11 | 46.8 | 46.8 | 48.1 | 47.7 | 33.7 | 44.6 | — | — | |
| Mix.AttackTarget VLM=LLAVA-7B [29]2024.11 | 46.8 | 47.6 | 47.6 | 48.2 | 34.3 | 44.9 | 1.7 | 3 | |
| MF-itTarget VLM=LLAVA-7B [29]2024.11 | 46.8 | 46.9 | 48 | 47.9 | 33.9 | 44.7 | 3 | 5.6 | |
| Mix.AttackTarget VLM=LLAVA-13B [29]2024.11 | 46.8 | 47.6 | 47.6 | 48.2 | 34.3 | 44.9 | 0.9 | 1.5 | |
| Clean imageTarget VLM=VIECAP [19]2024.11 | 46.7 | 44.3 | 47.7 | 47.2 | 35.2 | 44.2 | — | — | |
| MF-itTarget VLM=LLAVA-13B [29]2024.11 | 46.6 | 46.8 | 48 | 47.8 | 33.7 | 44.6 | 2.7 | 5 | |
| Clean imageTarget VLM=LLAVA-13B [29]2024.11 | 46.4 | 46.3 | 47.9 | 47.5 | 33.4 | 44.3 | — | — | |
| Mix.AttackTarget VLM=UNIDIFFUSER [4]2024.11 | 45.3 | 44 | 47.2 | 49.2 | 35.2 | 44.2 | 5.9 | 10.5 | |
| Clean imageTarget VLM=UNIDIFFUSER [4]2024.11 | 41.7 | 41.5 | 42.9 | 44.6 | 30.5 | 40.2 | — | — |