Image Captioning on Flickr30k (test) (L-infinity Metrics)
0.775Clean ScoreNo defense
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| No defenseBackbone=LLaVA 1.5-7B2025.06 | 0.775 | 0.016 | 0.01 | |
| CLIPureBackbone=LLaVA 1.5-7B2025.06 | 0.769 | 0.021 | 0.015 | |
| DiffCAPBackbone=LLaVA 1.5-7B2025.06 | 0.75 | 0.727 | 0.721 | |
| JPEG-DLBackbone=LLaVA 1.5-7B2025.06 | 0.748 | 0.696 | 0.479 | |
| FARE2Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 0.711 | 0.295 | 0.175 | |
| DiffPureBackbone=LLaVA 1.5-7B2025.06 | 0.655 | 0.664 | 0.632 | |
| FARE4Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 0.616 | 0.314 | 0.228 | |
| No defenseBackbone=OF-9B2025.06 | 0.601 | 0.007 | 0.004 | |
| CLIPureBackbone=OF-9B2025.06 | 0.593 | 0.047 | 0.035 | |
| JPEG-DLBackbone=OF-9B2025.06 | 0.588 | 0.476 | 0.307 | |
| FARE2Backbone=OF-9B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 0.577 | 0.164 | 0.089 | |
| TeCoA2Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 0.571 | 0.232 | 0.153 | |
| DiffCAPBackbone=OF-9B2025.06 | 0.556 | 0.567 | 0.572 | |
| FARE4Backbone=OF-9B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 0.514 | 0.157 | 0.105 | |
| DiffPureBackbone=OF-9B2025.06 | 0.498 | 0.492 | 0.503 | |
| TeCoA2Backbone=OF-9B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 0.495 | 0.141 | 0.095 | |
| TeCoA4Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 0.486 | 0.279 | 0.195 | |
| TeCoA4Backbone=OF-9B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 0.409 | 0.12 | 0.103 |