Indirect Prompt Injection Defense on Image Modality (test)
24.5UIAinjectIgnore Prompt
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Ignore Prompt2025.12 | 24.5 | 49.4 | 31.5 | 34.3 | 2 | |
| No Defense2025.12 | 30.9 | 49.6 | 25.1 | 26.8 | 0 | |
| Noise2025.12 | 34.3 | 46.8 | 7.6 | 10 | 1 | |
| System Prompt2025.12 | 38.2 | 42.7 | 10.7 | 11.4 | 6 | |
| AT2025.12 | 41.1 | 40.7 | 2.3 | 2.4 | 0 | |
| ARGUSsearch=disabled2025.12 | 44.5 | 49.6 | 0.1 | 0.1 | 3 | |
| ARGUSadaptive intervention=disabled2025.12 | 45.9 | 49.6 | 0.7 | 0.8 | 2 | |
| ARGUSvariant=full2025.12 | 46.3 | 49.6 | 0.1 | 0.1 | 3 | |
| ARGUSpost-filtering=disabled2025.12 | 46.4 | 49.6 | 4.3 | 4.8 | 3 | |
| Removal2025.12 | 48.5 | 49.3 | 0 | 0 | 12,885 |