Image Classification on MNIST (test) (Adversarial Robustness: PGD, CW, AA)
99.46Accuracy (Natural)Oracle
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Oracle2022.11 | 99.46 | 98.14 | 97.45 | 92.53 | — | |
| Warm-up+PLA2022.11 | 99.22 | 97.73 | 97.11 | 92.37 | — | |
| Two-stage2022.11 | 99.07 | 97.44 | 96.72 | 92.06 | — | |
| Clean TrainingAdversarial Perturbation=None2025.12 | 98.98 | — | — | — | 0 | |
| Selective Adversarial Training (Grad Matching)Sampling Strategy=Grad Matching, ρ (subset ratio)=0.252025.12 | 98.26 | — | — | — | 90.77 | |
| Selective Adversarial Training (Margin-based)Sampling Strategy=Margin-based, ρ (subset ratio)=0.252025.12 | 98.11 | — | — | — | 91.25 | |
| Full PGD-ATSampling Strategy=Full Minibatch, Perturbation=Multi-step PGD2025.12 | 97.78 | — | — | — | 91.14 | |
| Random Subset PGDSampling Strategy=Uniform Random, ρ (subset ratio)=0.252025.12 | 97.37 | — | — | — | 88.98 | |
| FORWARD2022.11 | 97.22 | 93.76 | 92.13 | 85.41 | — | |
| LOG2022.11 | 97.16 | 93.38 | 91.67 | 84.88 | — | |
| SCL_NL2022.11 | 93.09 | 87.07 | 84.59 | 75.98 | — | |
| NN2022.11 | 68.48 | 66.8 | 66.25 | 60.65 | — | |
| FREE2022.11 | 48.94 | 38.02 | 32.81 | 22.68 | — | |
| SCL_EXP2022.11 | 14.88 | 14.34 | 13.58 | 10.47 | — | |
| EXP2022.11 | 10.99 | 10.99 | 10.99 | 10.99 | — |