Image Classification on PCAM (Clean & AutoAttack Metrics)
58.26Clean AccuracyCLIP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CLIPBackbone=ViT-B/32, Evaluation Protocol=Zero-shot2026.03 | 58.26 | — | |
| HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.01 | 58.14 | 58.15 | |
| PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.01 | 57.7 | 57.65 | |
| AGFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 55.52 | — | |
| FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.01 | 54.74 | 48.15 | |
| PMG-AFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 54.22 | — | |
| CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.01 | 53.52 | 51.61 | |
| TeCoABackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 53.44 | — | |
| CLIPYear=20212026.04 | 52.26 | — | |
| CLIPFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 52.08 | 0 | |
| OursModel Configuration=5 trees2026.04 | 51.69 | — | |
| UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 51.2 | 29.51 | |
| OursModel Configuration=1 tree2026.04 | 50.9 | — | |
| AoSYear=20252026.04 | 50.22 | — | |
| ACATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 50.01 | 49.89 | |
| TRADESFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 49.95 | 46.42 | |
| GLADIATORBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 49.37 | — | |
| DKLFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 48.97 | 43.48 | |
| PMG-FTYear=20242026.04 | 48.47 | — | |
| TeCoAYear=20232026.04 | 47.23 | — | |
| TGA-ZSRBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 46.48 | — | |
| FAREYear=20242026.04 | 44.89 | — | |
| FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.01 | 37.89 | 53.68 |