Image Classification on CIFAR100 (Robust and Standard Accuracy)
88.8AccuracyViT-B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ViT-BAdaptation Strategy=Fine-tuning, Params=85.8M, Pre-training Dataset=ImageNet-1K2026.05 | 88.8 | — | |
| bViT-BAdaptation Strategy=Fine-tuning, Params=7.8M, Pre-training Dataset=ImageNet-1K2026.05 | 88 | — | |
| ViT-BAdaptation Strategy=LoRA, FFN, Params=738K, Pre-training Dataset=ImageNet-1K2026.05 | 87 | — | |
| ViT-BAdaptation Strategy=LoRA, QV, Params=296K, Pre-training Dataset=ImageNet-1K2026.05 | 86.6 | — | |
| bViT-BAdaptation Strategy=LoRA, FFN, Params=62K, Pre-training Dataset=ImageNet-1K2026.05 | 85.9 | — | |
| bViT-BAdaptation Strategy=LoRA, QV, Params=25K, Pre-training Dataset=ImageNet-1K2026.05 | 85.4 | — | |
| bViT-BAdaptation Strategy=Time embedding tuning, Params=9K, Pre-training Dataset=ImageNet-1K2026.05 | 83.9 | — | |
| ViT-BAdaptation Strategy=Linear probing, Params=—, Pre-training Dataset=ImageNet-1K2026.05 | 78.8 | — | |
| bViT-BAdaptation Strategy=Linear probing, Params=—, Pre-training Dataset=ImageNet-1K2026.05 | 77.2 | — | |
| No DefenseAttack Type=AutoAttack, Attack Budget (epsilon)=1/2552025.11 | 57.2 | 0.11 | |
| No DefenseAttack Type=AutoAttack, Attack Budget (epsilon)=4/2552025.11 | 57.2 | 0.11 | |
| No DefenseAttack Type=Carlini–Wagner, Attack Budget (epsilon)=1/2552025.11 | 57.2 | 0.3 | |
| No DefenseAttack Type=Carlini–Wagner, Attack Budget (epsilon)=4/2552025.11 | 57.2 | 0 | |
| ATACAttack Type=AutoAttack, Attack Budget (epsilon)=1/2552025.11 | 53.74 | 56.77 | |
| ATACAttack Type=AutoAttack, Attack Budget (epsilon)=4/2552025.11 | 53.74 | 57.49 | |
| ATACAttack Type=Carlini–Wagner, Attack Budget (epsilon)=1/2552025.11 | 53.74 | 53.75 | |
| ATACAttack Type=Carlini–Wagner, Attack Budget (epsilon)=4/2552025.11 | 53.74 | 78.08 |