Image Classification on SUN397 (Adversarial Robustness - Clean/AutoAttack)
54.26AutoAttack RobustnessCLBP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CLBPmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 54.26 | 65.24 | |
| SCCmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 48.99 | 58.25 | |
| TTCmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 39.44 | 55.13 | |
| HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.01 | 24.56 | 58.79 | |
| PMG-AFTmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 20.39 | 37.98 | |
| FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.01 | 19.21 | 47.01 | |
| TeCoAmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 18.36 | 36.69 | |
| UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 18.26 | 41.06 | |
| DKLFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 17.32 | 30.11 | |
| FAREmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 15.73 | 52.42 | |
| TRADESFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 15.69 | 26.64 | |
| ACATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 14.99 | 30.06 | |
| PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.01 | 11.13 | 58.78 | |
| CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.01 | 10.49 | 57.79 | |
| Anti-advmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 8.85 | 56 | |
| FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.01 | 5.71 | 46.61 | |
| CLIPmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 1.75 | 58.5 | |
| CLIP-FTmode=zero-shot, attack=CW-ℓ∞, epsilon_a=1/2552026.05 | 1.48 | 55.73 | |
| CLIPFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 0.7 | 57.26 |