Image Classification on Caltech 256 (Clean and Robust Accuracy)
90.5Clean AccuracyR-TPT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| R-TPTBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 90.5 | 85.4 | |
| TTEBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 89.9 | 71.3 | |
| CLIPBackbone=CLIP-L/14, Method category=Original2026.01 | 88.3 | 3.7 | |
| Anti-AdvBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 88 | 74.3 | |
| LPFBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 87.4 | 77.9 | |
| CSRBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 87.3 | 83.2 | |
| HDBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 86.2 | 39.5 | |
| FAREBackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 83.9 | 82.8 | |
| CLIPBackbone=ViT-B/32, Evaluation Protocol=Zero-shot2026.03 | 82.81 | — | |
| CLIPYear=20212026.04 | 82.01 | — | |
| TTCBackbone=CLIP-L/14, Method category=Test-Time Defense2026.01 | 82 | 44 | |
| OursModel Configuration=5 trees2026.04 | 81.85 | — | |
| OursModel Configuration=1 tree2026.04 | 81.27 | — | |
| AoSYear=20252026.04 | 80.93 | — | |
| R-TPTBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 80.8 | 63.5 | |
| CLIPBackbone=CLIP-B/32, Category=Original2026.01 | 80.3 | 4.1 | |
| AGFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 80.26 | — | |
| GLADIATORBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 79.44 | — | |
| Anti-AdvBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 79.2 | — | |
| CSRBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 79.2 | 55 | |
| TeCoABackbone=CLIP-L/14, Method category=Adversarial Fine-tuning2026.01 | 79.2 | 78.7 | |
| FAREYear=20242026.04 | 79.13 | — | |
| HDBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 78.8 | 21.7 | |
| LPFBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 78.6 | 49.3 | |
| TTEBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 78.5 | 60 | |
| PMG-AFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 78.33 | — | |
| TGA-ZSRBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 77.45 | — | |
| PMG-FTYear=20242026.04 | 77.4 | — | |
| TeCoABackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 77.34 | — | |
| TTCBackbone=CLIP-B/32, Category=Test-Time Defense2026.01 | 77.3 | 49.2 | |
| FAREBackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 75.6 | 72.8 | |
| TeCoAYear=20232026.04 | 73.51 | — | |
| TeCoABackbone=CLIP-B/32, Category=Adversarial Fine-tuning2026.01 | 66.6 | 64.1 | |
| AGFTFine-tuned on=ImageNet2026.03 | — | 67.95 | |
| AoSZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 63.98 | |
| CLIPMode=Zero-shot2026.03 | — | 11.69 | |
| CLIPZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 14 | |
| FAREZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 58.63 | |
| GLADIATORFine-tuned on=ImageNet2026.03 | — | 63.5 | |
| OursZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 64.15 | |
| Ours (5 trees)Zero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/255, Number of trees=52026.04 | — | 65.41 | |
| PMG-AFTFine-tuned on=ImageNet2026.03 | — | 63.15 | |
| PMG-FTZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 59.57 | |
| TeCoAFine-tuned on=ImageNet2026.03 | — | 61.23 | |
| TeCoAZero-shot=true, Adversarial fine-tuning=ImageNet, PGD attack steps=20, Perturbation radius (epsilon)=1/2552026.04 | — | 59.8 | |
| TGA-ZSRFine-tuned on=ImageNet2026.03 | — | 64.06 |