Image Classification Robustness on Cifar10
88.03Clean AccuracyCLIP
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CLIPTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 88.03 | 9.57 | — | 2.54 | |
| CLIP2025.12 | 88.03 | 2.54 | 9.91 | 9.57 | |
| CLIPTraining Dataset=ImageNet-21K, Evaluation Protocol=Zero-shot2025.12 | 88.03 | 0.05 | 0.58 | 0.94 | |
| CLIPFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 88.02 | 6.47 | — | — | |
| FARETraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 84.75 | 50.96 | — | 49.62 | |
| TGA-ZSRTraining Dataset=TinyImageNet, Adversarial Training=10-step PGD, Training Epsilon=2/255, Evaluation Protocol=Zero-shot2025.12 | 83.98 | 6.29 | 20.48 | 20.17 | |
| UCATTraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 83.78 | 50.59 | — | 49 | |
| UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 81.81 | 42.32 | — | — | |
| UCATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 81.81 | 42.32 | 43.3 | 43.81 | |
| FAREFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 79.81 | 0.03 | 4.12 | 7.83 | |
| TGA-ZSRFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 79.18 | 40.92 | 42.24 | 42.16 | |
| TeCoATraining Data=ImageNet-1k, Adversarial Training=2-step PGD, Epsilon=1/2552025.12 | 78.12 | 59.28 | — | 58.27 | |
| PMG-AFTFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 74.73 | 41.86 | 42.1 | 51.39 | |
| PMG-AFTTraining Dataset=TinyImageNet, Adversarial Training=10-step PGD, Training Epsilon=2/255, Evaluation Protocol=Zero-shot2025.12 | 74.05 | 0.96 | 1.63 | 24.1 | |
| TeCoAFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/2552025.12 | 67.56 | 37.87 | 38.62 | 39.33 | |
| UCATTraining Dataset=TinyImageNet, Adversarial Training=10-step PGD, Training Epsilon=2/255, Evaluation Protocol=Zero-shot2025.12 | 66.52 | 24.08 | 25.46 | 25.81 | |
| DKLFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 65.31 | 41.77 | — | — | |
| ACATFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 64.72 | 34.54 | — | — | |
| TRADESFine-tuning dataset=TinyImageNet, Adversarial training=2-step PGD, Epsilon=1/255, Evaluation protocol=Zero-shot2025.12 | 62.05 | 41.39 | — | — | |
| TeCoATraining Dataset=TinyImageNet, Adversarial Training=10-step PGD, Training Epsilon=2/255, Evaluation Protocol=Zero-shot2025.12 | 58.62 | 21.94 | 23.05 | 23.57 | |
| FARETraining Dataset=TinyImageNet, Adversarial Training=10-step PGD, Training Epsilon=2/255, Evaluation Protocol=Zero-shot2025.12 | 40.23 | 14.81 | 14.64 | 18.19 |