Certified Robust Image Classification on CIFAR-10 (test) (Arob Metrics)
85.35Anat AccuracyHYDRA
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HYDRABackbone=WRN-28-4, Training Mode=Adversarial Training, Size [MB]=22.3, FPS=172021.10 | 85.35 | — | — | — | — | 57.23 | |
| HYDRA + GDWSBackbone=WRN-28-4, Training Mode=Adversarial Training, Beta (β)=1, Size [MB]=20.5, FPS=682021.10 | 84.17 | — | — | — | — | 55.87 | |
| HYDRABackbone=WRN-28-4, Pruning Ratio (p)=90%, Size [MB]=2.3, FPS=172021.10 | 83.69 | — | — | — | — | 55.2 | |
| HYDRA + GDWSBackbone=WRN-28-4, Pruning Ratio (p)=90%, Beta (β)=0.125, Size [MB]=11.9, FPS=592021.10 | 83.38 | — | — | — | — | 54.79 | |
| HYDRABackbone=VGG-16, Training Mode=Adversarial Training, Size [MB]=58.4, FPS=362021.10 | 82.72 | — | — | — | — | 51.93 | |
| HYDRABackbone=WRN-28-4, Pruning Ratio (p)=95%, Size [MB]=1.1, FPS=172021.10 | 82.68 | — | — | — | — | 54.18 | |
| HYDRA + GDWSBackbone=WRN-28-4, Pruning Ratio (p)=95%, Beta (β)=0.005, Size [MB]=7.2, FPS=602021.10 | 82.59 | — | — | — | — | 54.22 | |
| HYDRA + GDWSBackbone=VGG-16, Training Mode=Adversarial Training, Beta (β)=0.5, Size [MB]=50.6, FPS=1022021.10 | 82.53 | — | — | — | — | 50.96 | |
| ResNet-18 (AT)Adversarial Training Source=[23], FPS=282021.10 | 81.74 | 47.5 | 53.6 | 66.1 | 46.1 | — | |
| ResNet-18 (AT) + GDWSAdversarial Training Source=[23], beta=0.0025, FPS=872021.10 | 81.67 | 47.6 | 53.6 | 66 | 46.3 | — | |
| ResNet-18 (AT) + GDWSAdversarial Training Source=[23], beta=0.005, FPS=922021.10 | 81.43 | 47.3 | 52.6 | 65.6 | 45.7 | — | |
| ResNet-18 (AT) + GDWSAdversarial Training Source=[23], beta=0.01, FPS=1012021.10 | 81.1 | 47.2 | 52.2 | 65 | 45.2 | — | |
| HYDRABackbone=VGG-16, Pruning Ratio (p)=90%, Size [MB]=5.9, FPS=362021.10 | 80.54 | — | — | — | — | 49.44 | |
| HYDRA + GDWSBackbone=VGG-16, Pruning Ratio (p)=90%, Beta (β)=0.1, Size [MB]=31.5, FPS=932021.10 | 80.47 | — | — | — | — | 49.52 | |
| HYDRABackbone=VGG-16, Pruning Ratio (p)=95%, Size [MB]=3, FPS=362021.10 | 78.91 | — | — | — | — | 48.74 | |
| HYDRA + GDWSBackbone=VGG-16, Pruning Ratio (p)=95%, Beta (β)=0.1, Size [MB]=18.3, FPS=1062021.10 | 78.71 | — | — | — | — | 48.53 | |
| HYDRABackbone=WRN-28-4, Pruning Ratio (p)=99%, Size [MB]=0.2, FPS=282021.10 | 75.62 | — | — | — | — | 47.21 | |
| HYDRA + GDWSBackbone=WRN-28-4, Pruning Ratio (p)=99%, Beta (β)=0.0025, Size [MB]=1.2, FPS=682021.10 | 75.36 | — | — | — | — | 47.04 | |
| HYDRABackbone=VGG-16, Pruning Ratio (p)=99%, Size [MB]=0.6, FPS=412021.10 | 73.16 | — | — | — | — | 41.74 | |
| HYDRA + GDWSBackbone=VGG-16, Pruning Ratio (p)=99%, Beta (β)=0.02, Size [MB]=2.9, FPS=1362021.10 | 72.75 | — | — | — | — | 41.56 |