Adversarial Robustness (Epsilon 2 & 3) on MNIST (Image Classification)
98.6Standard AccuracyPGD adversarial training
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| PGD adversarial trainingAttack Type=Square Attack, Queries=20002024.08 | 98.6 | 25.6 | 1.2 | |
| l∞-dist netAttack Type=Square Attack, Queries=20002024.08 | 98.5 | 14 | 16.3 | |
| FGSM adversarial trainingAttack Type=Square Attack, Queries=20002024.08 | 97.7 | 0 | 0 | |
| Median filterAttack Type=Square Attack, Queries=20002024.08 | 97.6 | 0 | 0 | |
| STLAttack Type=Square Attack, Queries=20002024.08 | 96.4 | 0 | 0 | |
| OTADAttack Type=Square Attack, Queries=20002024.08 | 96.3 | 63 | 41.6 | |
| SOC+Attack Type=Square Attack, Queries=20002024.08 | 96 | 56.4 | 26.1 | |
| APE-GANAttack Type=Square Attack, Queries=20002024.08 | 93.3 | 4.2 | 0.2 |