Image Classification on ImageNet-S (Robustness Evaluation)
51.1AccuracyDBD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DBDAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 51.1 | 95.3 | |
| CLIPAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 49.1 | 0 | |
| R-TPTAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 47.9 | 30.3 | |
| DBDAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 45.1 | 89.5 | |
| Vanilla FTBackbone=ViT-B/322026.03 | 43.34 | 0 | |
| WiSE-FTBackbone=ViT-B/322026.03 | 43.34 | 0 | |
| TPGMBackbone=ViT-B/322026.03 | 42.71 | 0 | |
| SPDBackbone=ViT-B/322026.03 | 42.56 | 0 | |
| CLIPAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 42.5 | 0 | |
| FLYPBackbone=ViT-B/322026.03 | 42.22 | 0 | |
| GRACEBackbone=ViT-B/322026.03 | 42.11 | 17.75 | |
| MTACategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack scenario=PGD-100 (eps=4.0)2026.06 | 41.8 | 9.4 | |
| R-TPTAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 41.7 | 23.6 | |
| CLIPBackbone=ViT-B/322026.03 | 41.44 | 0 | |
| R-TPTCategory=Tuning-based, Backbone=CLIP-ViT-B/32, Attack scenario=PGD-100 (eps=4.0)2026.06 | 40.4 | 22.7 | |
| CLIPCategory=Baseline, Backbone=CLIP-ViT-B/32, Attack scenario=PGD-100 (eps=4.0)2026.06 | 39.8 | 0 | |
| MACCategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack scenario=PGD-100 (eps=4.0)2026.06 | 39.6 | 33.5 | |
| AGFTAttack protocol=PGD-20, Perturbation budget ε=1/2552026.03 | 39.58 | 29.63 | |
| TGA-ZSRAttack protocol=PGD-20, Perturbation budget ε=1/2552026.03 | 38.5 | 27.49 | |
| TTCAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/162026.06 | 38.5 | 1.9 | |
| DBDAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 38.1 | 73.7 | |
| CLIPAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 35.6 | 0.3 | |
| PMG-AFTAttack protocol=PGD-20, Perturbation budget ε=1/2552026.03 | 35.38 | 24.07 | |
| TTCCategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack scenario=PGD-100 (eps=4.0)2026.06 | 35.3 | 5.4 | |
| MTABackbone=CLIP-ResNet502025.04 | 35.2 | 11.3 | |
| TPTBackbone=CLIP-ResNet502025.04 | 35 | 1.4 | |
| C-TPTBackbone=CLIP-ResNet502025.04 | 34.7 | 0.9 | |
| TPT2026.06 | 34.7 | 1.4 | |
| TeCoAAttack protocol=PGD-20, Perturbation budget ε=1/2552026.03 | 34.41 | 22.92 | |
| R-TPTBackbone=CLIP-ResNet502025.04 | 34 | 26.2 | |
| R-TPTAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 34 | 26 | |
| CLIPBackbone=CLIP-ResNet502025.04 | 33.3 | 0.5 | |
| CLIP2026.06 | 33.3 | 0.7 | |
| SS-TPT2026.06 | 33.2 | 25.4 | |
| R-TPT2026.06 | 32.4 | 21.3 | |
| PMG-AFTBackbone=ViT-B/322026.03 | 31.1 | 14.8 | |
| DOC2026.06 | 30.8 | 14.4 | |
| TTC2026.06 | 30.6 | 14.5 | |
| LAATBackbone=ViT-B/322026.03 | 30.15 | 12.2 | |
| TTCAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 29.8 | 12.8 | |
| Ensemble2026.06 | 29.6 | 17.4 | |
| EnsembleBackbone=CLIP-ResNet502025.04 | 29.5 | 20.7 | |
| FAREBackbone=ViT-B/322026.03 | 28.63 | 11.05 | |
| TeCoABackbone=ViT-B/322026.03 | 25.89 | 13.06 | |
| TeCoAAttack=PGD-100, Epsilon=4/255, Backbone=CLIP-ViT-B/322026.06 | 18.5 | 7.6 | |
| TeCoAAttack=PGD-10, Epsilon=1/255, Backbone=CLIP-ResNet502026.06 | 18.3 | 11.6 |