Image Classification on SVHN (Standard and Robust Accuracy via AutoAttack)
95.39Accuracy (Standard)RAMRM-PureMaskDiT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RAMRM-PureMaskDiTExtra data=✗, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 95.39 | — | 66.12 | |
| Wang et al.Extra data=✓, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 95.19 | — | 61.85 | |
| AMRM-PureMaskDiTExtra data=✗, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 94.91 | — | 64.15 | |
| AMRM-PureMAEExtra data=✗, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 94.54 | — | 47.03 | |
| RAMRM-PureMAEExtra data=✗, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 94.47 | — | 50.03 | |
| Rebuffi et al.Extra data=✓, Architecture=WRN-28-10, Attack Budget (epsilon)=8/2552026.07 | 94.34 | — | 60.9 | |
| WideResNet-50Setup=Adversarial Fine-tuning (Tad), Compression=INT8 Quantization2024.03 | 90.87 | 38.76 | — | |
| WideResNet-50Setup=Standard Baseline (fst), Compression=None2024.03 | 90.42 | 6.51 | — | |
| WideResNet-50Setup=Robust Baseline (frb), Compression=None2024.03 | 89.51 | 38.37 | — | |
| WideResNet-50Setup=Standard Fine-tuning (Tst), Compression=INT8 Quantization2024.03 | 89.33 | 2.55 | — | |
| WideResNet-50Setup=Adversarial Fine-tuning (Tad), Compression=l1-Pruning (80% sparsity)2024.03 | 88.29 | 35.46 | — | |
| WideResNet-50Setup=Standard Fine-tuning (Tst), Compression=l1-Pruning (80% sparsity)2024.03 | 88.2 | 9.47 | — | |
| ViTSetup=Standard Fine-tuning (Tst), Compression=INT8 Quantization2024.03 | 84.97 | 0.29 | — | |
| ViTSetup=Standard Baseline (fst), Compression=None2024.03 | 84.43 | 0.39 | — | |
| ViTSetup=Standard Fine-tuning (Tst), Compression=l1-Pruning (80% sparsity)2024.03 | 82.14 | 1.02 | — | |
| ViTSetup=Adversarial Fine-tuning (Tad), Compression=INT8 Quantization2024.03 | 81.5 | 27.45 | — | |
| ViTSetup=Robust Baseline (frb), Compression=None2024.03 | 79.59 | 21.38 | — | |
| ViTSetup=Adversarial Fine-tuning (Tad), Compression=l1-Pruning (80% sparsity)2024.03 | 77.22 | 17.42 | — |