Certified Robustness on CIFAR-10 l2 (test)
768,560Average Sample CountHorvath
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Horvathepsilon (ε)=0.01, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 768,560 | — | |
| Betting-CSepsilon (ε)=0.01, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 199,771 | — | |
| UB-CSepsilon (ε)=0.01, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 197,628 | — | |
| Horvathepsilon (ε)=0.02, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 94,900 | — | |
| Horvathepsilon (ε)=0.03, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 81,080 | — | |
| UB-CSepsilon (ε)=0.02, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 49,198 | — | |
| Betting-CSepsilon (ε)=0.02, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 47,215 | — | |
| UB-CSepsilon (ε)=0.03, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 21,513 | — | |
| Betting-CSepsilon (ε)=0.03, Backbone=WideResNet-40, sigma=1, confidence level=0.9992024.06 | 20,918 | — | |
| Adaptive Horváth et al. (2022)radius (r)=2, Backbone=WideResNet-40, sigma=12024.06 | 4,623 | 0.3 | |
| Adaptive Horváth et al. (2022)radius (r)=1.25, Backbone=WideResNet-40, sigma=12024.06 | 3,593 | 0.23 | |
| Union bound CS 1radius (r)=2, Backbone=WideResNet-40, sigma=12024.06 | 2,670 | 0.21 | |
| Union bound CS 1radius (r)=1.25, Backbone=WideResNet-40, sigma=12024.06 | 2,557 | 0.19 | |
| Betting CS 2radius (r)=1.25, Backbone=WideResNet-40, sigma=12024.06 | 2,169 | 0.17 | |
| Betting CS 2radius (r)=2, Backbone=WideResNet-40, sigma=12024.06 | 2,130 | 0.17 | |
| Adaptive Horváth et al. (2022)radius (r)=0.5, Backbone=WideResNet-40, sigma=12024.06 | 1,976 | 0.13 | |
| Union bound CS 1radius (r)=0.5, Backbone=WideResNet-40, sigma=12024.06 | 635 | 0.05 | |
| Betting CS 2radius (r)=0.5, Backbone=WideResNet-40, sigma=12024.06 | 531 | 0.05 |