Image Classification on SUN397 (Adversarial Robustness)
59.01Clean AccuracyCLIP
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| CLIPBackbone=ViT-B/32, Evaluation Protocol=Zero-shot2026.03 | 59.01 | — | — | |
| CLIPTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 57.26 | 1.02 | 0.05 | |
| AGFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 55.19 | — | — | |
| TeCoABackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 54.45 | — | — | |
| UCATTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 53.98 | 29.93 | 27.85 | |
| FARETraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 53.97 | 26.66 | 24.77 | |
| PMG-AFTBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 53.96 | — | — | |
| GLADIATORBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 53.61 | — | — | |
| TeCoATraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 51.28 | 29.81 | 29.03 | |
| TGA-ZSRBackbone=ViT-B/32, Evaluation Protocol=Zero-shot, Training Strategy=Adversarially fine-tuned2026.03 | 50.1 | — | — |