List your company
ResearchFormerly WizWand

Adversarial Robustness Evaluation on FGVC Aircraft (Image Classification)

71.9Robust AccuracyDBD
-2.87616.53735.9555.363Feb 12, 2026Mar 5, 2026Mar 27, 2026Apr 18, 2026May 9, 2026May 31, 2026Jun 22, 2026
Updated 4mo ago

Evaluation results

MethodLinks
DBDBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.0671.929.7———
DBDBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.0666.322.4———
DBDBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.0656.321.6———
R-TPTBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.0614.424———
R-TPTBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.0612.318.1———
TTCBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.0611.115.4———
R-TPTBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.0610.119.6———
APTBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Few-shot2026.066.828.7———
Comp-TGAProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.035.41————
TGA-ZSRProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.034.75————
Comp-TGAMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.034.56————
SAFT-LZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.024.5313.39———
TGA-ZSRMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.034.26————
SAFT-MZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.024.0213.86———
Comp-TGAZero-shot=true, Attack=APGD (epsilon=1/255)2026.033.97————
TGA-ZSRZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.023.5111.98———
TGA-ZSRZero-shot=true, Attack=APGD (epsilon=1/255)2026.033.12————
T-VSSBackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.062.97.3———
TTCBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.062.821.7———
R-TPTBackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.062.76.1———
TeCoABackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.062.46.6———
EnsembleBackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.0624.2———
PMG-AFTProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.98————
Comp-TGAZero-shot=true2026.031.95————
TeCoAZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.021.876.42———
PMG-AFTZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.021.853.84———
FT-Adv.Protocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.83————
FT-Adv.Zero-shot=true, Attack=APGD (epsilon=1/255)2026.031.8————
PMG-AFTZero-shot=true, Attack=APGD (epsilon=1/255)2026.031.71————
TPTBackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.061.76.7———
TeCoAZero-shot=true, Attack=APGD (epsilon=1/255)2026.031.68————
TeCoAMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.62————
LAATMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.62————
TeCoAProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.62————
TGA-ZSRZero-shot=true2026.031.61————
MTABackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.061.66.5———
PMG-AFTMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.56————
LAATZero-shot=true, Attack=APGD (epsilon=1/255)2026.031.53————
LAATProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.38————
FT-Adv.Mode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.031.26————
FAREZero-shot=true, Attack=APGD (epsilon=1/255)2026.031.08————
PMG-AFTZero-shot=true2026.030.9————
FAREZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.020.8414.19———
FAREMode=Zero-shot, Attack=Adaptive Attack, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.030.84————
FAREProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.030.84————
FT-Adv.Zero-shot=true2026.030.83————
C-TPTBackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.060.76.4———
TeCoAZero-shot=true2026.030.63————
LAATZero-shot=true2026.030.63————
TeCoABackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.060.66.7———
CLIP-TeCoABackbone=TeCoA-pretrained CLIP-ViT-B/32, Epsilon (ϵ)=4/2552026.060.56.6———
FAREZero-shot=true2026.030.28————
CLIPZero-shot=true, Attack=APGD (epsilon=1/255)2026.030.06————
CLIPZero-shot evaluation=true, Adversarial Attack=PGD-100 (epsilon = 1/255), Backbone=ViT-B/322026.020.039.3———
FT-CleanProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.030.03————
FT-CleanZero-shot=true, Attack=APGD (epsilon=1/255)2026.030.03————
FT-CleanZero-shot=true2026.030.01————
CLIPZero-shot=true2026.030————
CLIPProtocol=Zero-shot, Attack=PGD, Epsilon=1/255, Training Dataset=Tiny-ImageNet2026.030————
CLIPBackbone=ResNet50, Attack Protocol=PGD-10, Epsilon (ε)=1/255, Evaluation Setting=Zero-shot2026.06019.6———
CLIPBackbone=ViT-B/32, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06020.6———
CLIPBackbone=ViT-B/16, Attack Protocol=PGD-100, Epsilon (ε)=4/255, Evaluation Setting=Zero-shot2026.06027.1———
ACATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12—4.411.59——
CLIPFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12—20.160——
CLIPTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12—20.1600.23—
CLIP2025.12—20.160.2300
DKLFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12—3.961.65——
FARETraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12—17.15.284.08—
FAREFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12—15.630.060.541.11
PMG-AFTFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12—11.790.513.210.99
TeCoATraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12—13.895.14.05—
TeCoAFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12—5.191.862.162.1
TGA-ZSRFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12—8.822.342.882.85
TRADESFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12—3.61.23——
UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12—10.952.61——
UCATTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12—15.245.734.5—
UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12—10.952.613.453.21