Fine-grained classification on FGVC Aircraft
88.89AccuracyResNet-50 (ReLabel-trained)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ResNet-50 (ReLabel-trained)Backbone=ResNet-50, Pre-training strategy=ReLabel2021.01 | 88.89 | — | — | |
| ResNet-50 (Baseline)Backbone=ResNet-50, Pre-training strategy=Standard2021.01 | 85.09 | — | — | |
| MTAPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 32.7 | 8 | — | |
| EnsemblePre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 31.7 | 12.7 | — | |
| R-TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 31.7 | 17.2 | — | |
| TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 31.6 | 0 | — | |
| C-TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 30.4 | 0 | — | |
| CLIPPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 30 | 0 | — | |
| Clean (Reference)Backbone=CLIP-ViT-B/16, Labeling Strategy=none (clean accuracy)2026.06 | 27.2 | — | — | |
| MTABackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 25 | — | 4.3 | |
| MTABackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 25 | 4.3 | — | |
| AGCBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 24.2 | 68.6 | — | |
| CLIPBackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 23.9 | — | 0 | |
| C-TPTBackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 23.9 | — | 0 | |
| R-TPTBackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 23.9 | — | 13.2 | |
| CLIPBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 23.9 | 0 | — | |
| R-TPTBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 23.9 | 13.2 | — | |
| TTPBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 23.9 | 14.8 | — | |
| EnsembleBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 23.5 | 12.7 | — | |
| EnsembleBackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 23.4 | — | 8.7 | |
| TPTBackbone=CLIP-ViT-B/16, Epsilon (ε)=4.02025.04 | 23.4 | — | 0 | |
| TTCBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 23.3 | 0.5 | — | |
| MTABackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 20.2 | 8.2 | — | |
| MTABackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 20.2 | 8.2 | — | |
| AGCBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 19.9 | 56.1 | — | |
| T-VSSBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 19.6 | 15.8 | — | |
| EnsembleBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 19.5 | 9.3 | — | |
| R-TPTPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 19.1 | 9.2 | — | |
| R-TPTBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 19.1 | 9.2 | — | |
| R-TPTBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 19.1 | 9.2 | — | |
| TTCBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 18.2 | 1 | — | |
| TTCBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 18.2 | 1 | — | |
| CLIPPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 18.1 | 0 | — | |
| CLIPBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 18.1 | 0 | — | |
| CLIPBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Baseline2026.06 | 18.1 | 0 | — | |
| TTPBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 18 | 10.3 | — | |
| TTPBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 18 | 10.3 | — | |
| R-TPTPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 17.5 | 12.6 | — | |
| EnsembleBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Test-time Defense2026.06 | 16.4 | 7.9 | — | |
| CLIPPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 15.7 | 0 | — | |
| APT+TeCoAPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 9.9 | 9.7 | — | |
| FAREPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 9.5 | 8.5 | — | |
| FAREBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Training-time Defense2026.06 | 9.5 | 8.5 | — | |
| TeCoAPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 6.6 | 5.9 | — | |
| TeCoABackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Training-time Defense2026.06 | 6.6 | 5.9 | — | |
| TeCoAPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 5.8 | 5.8 | — | |
| APT+TeCoAPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 5.2 | 5 | — | |
| APT+TeCoABackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Training-time Defense2026.06 | 5.2 | 5 | — | |
| APTPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 2.6 | 0.5 | — | |
| APTBackbone=CLIP-ViT-B/32, Epsilon (ϵ)=4/255, Defense Category=Training-time Defense2026.06 | 2.6 | 0.5 | — | |
| APTPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 0.9 | 0.9 | — | |
| CLIPBackbone=CLIP-ViT-B/16, Attack Type=PGD-100, Perturbation Budget (epsilon)=4/255, Labeling Strategy=pseudo-labels2026.06 | — | 1.5 | — | |
| DBDBackbone=CLIP-ViT-B/16, Attack Type=PGD-100, Perturbation Budget (epsilon)=4/255, Labeling Strategy=pseudo-labels2026.06 | — | 26.3 | — | |
| R-TPTBackbone=CLIP-ViT-B/16, Attack Type=PGD-100, Perturbation Budget (epsilon)=4/255, Labeling Strategy=pseudo-labels2026.06 | — | 18.2 | — |