Image Classification on Caltech101 (Clean and Robust Accuracy)
85.32Clean AccuracyCLIP
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CLIPTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 85.32 | — | — | 24.63 | 0.1 | — | |
| CLIP2025.12 | 85.32 | — | — | 0.1 | 5.65 | 6.82 | |
| FARETraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 85.01 | — | — | 68.19 | 67.21 | — | |
| UCATTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 84.64 | — | — | 71.11 | 70.12 | — | |
| TeCoATraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 83.51 | — | — | 69.07 | 68.4 | — | |
| PMG-AFTFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 82.49 | — | — | 44.91 | 61.13 | 52.84 | |
| UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 82.39 | — | — | 63.73 | 65.32 | 64.63 | |
| FAREFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 78.22 | — | — | 10.26 | 33.21 | 32.5 | |
| TGA-ZSRFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 76.42 | — | — | 57.11 | 58.05 | 57.72 | |
| TeCoAFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 72.16 | — | — | 54.95 | 56 | 55.62 | |
| FT-StandardBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=4062026.01 | 29.1 | — | 20.31 | — | — | — | |
| FT-TeCoABackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=11122026.01 | 28.32 | — | 24.02 | — | — | — | |
| HPT-GPDBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=14862026.01 | 22.87 | — | 19.11 | — | — | — | |
| PMGBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=18172026.01 | 20.98 | — | 14.82 | — | — | — | |
| CLIPBackbone=ViT-B/32, Proxy CLIP=ViT-B/16, Attack Protocol=PGD-10, Training time per epoch (s)=02026.01 | 19.42 | — | 15.82 | — | — | — |