Image Classification on Food101 (Clean and Robust Accuracy)
90.8Clean AccuracyCLIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CLIPBackbone=CLIP-L/14, Method category=Original2026.01 | 90.8 | 0.001 | — | |
| R-TPTBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 90.6 | 0.793 | — | |
| CSRBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 90.6 | 0.858 | — | |
| TTEBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 89.8 | 0.368 | — | |
| HDBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 88.7 | 0.07 | — | |
| Anti-AdvBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 87.7 | 0.531 | — | |
| LPFBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 86.8 | 0.612 | — | |
| CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.01 | 82.93 | — | 3.28 | |
| R-TPTBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 82.1 | 0.571 | — | |
| CLIPBackbone=CLIP-B/32, Category=Original2026.01 | 80.4 | 0.001 | — | |
| TTEBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 79.9 | 0.46 | — | |
| HDBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 79.7 | 0.036 | — | |
| CSRBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 79.4 | 0.577 | — | |
| Anti-AdvBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 78.5 | 0.114 | — | |
| HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.01 | 75.05 | — | 22.24 | |
| TTCBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 72.7 | 0.365 | — | |
| LPFBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 71 | 0.203 | — | |
| PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.01 | 70.52 | — | 13.91 | |
| FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.01 | 70.07 | — | 2.65 | |
| TTCBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 68.3 | 0.479 | — | |
| FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.01 | 52.81 | — | 14.45 | |
| FAREBackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 49.4 | 0.48 | — | |
| FAREBackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 38.3 | 0.366 | — | |
| TeCoABackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 36.2 | 0.348 | — | |
| TeCoABackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 24.1 | 0.221 | — | |
| CLIPEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01 | — | 0.0039 | — | |
| HPT-GPDEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01 | — | 0.0512 | — | |
| PMGEvaluation Protocol=Zero-shot, Attack Strategy=AutoAttack, Perturbation Budget=1/2552026.01 | — | 0.0044 | — |