Image Classification on MNIST (Certified Robustness Accuracy)
99.4Clean Accuracyl2
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| l2Perturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 99.4 | 0 | 94.5 | 94.7 | 0 | |
| l_infinityPerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98.7 | 92.1 | 69.6 | 38.9 | 38.5 | |
| CURE-JointPerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98.7 | 90.5 | 76.3 | 50.8 | 50.3 | |
| CURE-MaxPerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98.7 | 91.1 | 76.2 | 47.2 | 46.5 | |
| CURE-RandomPerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98.7 | 90.5 | 76.3 | 50.8 | 50.3 | |
| CURE-FinetunePerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98.5 | 90.1 | 83.5 | 64 | 63.2 | |
| CURE-ScratchPerturbation Budget (epsilon_inf, epsilon_2, epsilon_1)=(0.3, 1.0, 2.0)2024.10 | 98 | 89.4 | 85.9 | 71.5 | 70.5 |