caption_blip_base
not supported
- Distance ratio
- 2.437
- 95% prompt CI
- [2.101, 2.819]
- PCA ratio
- 2.785
- Paired log gain
- -0.029
- Paired 95% CI
- [-0.081, 0.022]
- Recall@1
- 0.394
- Triplet accuracy
- 0.667
Separate learned metrics for frozen DINO image embeddings and frozen caption embeddings. Conclusions are opened only on untouched prompts and held-out attack families.
large_scale_attack_invariant_v1
| Track | Group | Value | Ratio |
|---|---|---|---|
| dinov2_base | attack_family | occlusion_overlay | 23.726 |
| dinov2_base | attack_family | social_media_compound | 52.402 |
| dinov2_base | severity | medium | 38.608 |
| dinov2_base | severity | mild | 202.008 |
| dinov2_base | severity | severe | 12.977 |
| dinov2_base | prompt_source | coco2017 | 23.546 |
| dinov2_base | prompt_source | diffusiondb | 23.326 |
| dinov2_base | caption_divergence | high | 26.522 |
| dinov2_base | caption_divergence | low | 16.841 |
| dinov2_base | caption_divergence | medium | 26.944 |
| dinov3_l16 | attack_family | occlusion_overlay | 36.725 |
| dinov3_l16 | attack_family | social_media_compound | 20.299 |
| dinov3_l16 | severity | medium | 38.211 |
| dinov3_l16 | severity | mild | 134.723 |
| dinov3_l16 | severity | severe | 16.239 |
| dinov3_l16 | prompt_source | coco2017 | 14.926 |
| dinov3_l16 | prompt_source | diffusiondb | 21.697 |
| dinov3_l16 | caption_divergence | high | 24.385 |
| dinov3_l16 | caption_divergence | low | 12.882 |
| dinov3_l16 | caption_divergence | medium | 17.668 |
| caption_blip_base | attack_family | occlusion_overlay | 18674430583332.262 |
| caption_blip_base | attack_family | social_media_compound | 209386037637.075 |
| caption_blip_base | severity | medium | 33521335307699.848 |
| caption_blip_base | severity | mild | 57554702363810.000 |
| caption_blip_base | severity | severe | 13334414092827.037 |
| caption_blip_base | prompt_source | coco2017 | 2.493 |
| caption_blip_base | prompt_source | diffusiondb | 2.382 |
| caption_blip_base | caption_divergence | high | 2.766 |
| caption_blip_base | caption_divergence | low | 2.210 |
| caption_blip_base | caption_divergence | medium | 2.336 |









Exact collision rate: 0.741. Empty rate: 0.000.
Image/caption semantic-distance Spearman correlations: