Table 4 Maximum test accuracy obtained through regular training vs fine-tuning over different benchmark datasets with different competing techniques when no adversarial attack is performed
From: Gaussian class-conditional simplex loss for accurate, adversarially robust deep classifier training
Method | MNIST ResNet-18 | FMNIST ResNet-18 | SVHN ResNet-18 | CIFAR-10 ResNet-18 | CIFAR-10 Shake-Shake-96 | CIFAR-100 Shake-Shake-112 |
|---|---|---|---|---|---|---|
GCCS (regular training) | 99.58 | 92.69 | 94.20 | 82.97 | 96.19 | 76.53 |
GCCS (fine-tuning) | 99.64 | 93.83 | 95.58 | 81.52 | 97.06 | 77.48 |
No Defense (cross-entropy loss) | 99.35 | 91.91 | 94.12 | 78.59 | 95.78 | 76.30 |
Jacobian Reg. (regular training) [53] | 98.99 | 91.79 | 94.11 | 70.09 | - | - |
Jacobian Reg. (fine-tuning) [53] | 98.53 | 92.43 | 93.54 | 82.09 | - | - |
Input Gradient Reg. (regular training) [44] | 97.98 | 88.45 | 93.77 | 78.32 | 96.50 | 74.89 |
Input Gradient Reg. (fine-tuning) [44] | 99.11 | 92.55 | 93.17 | 76.15 | 96.90 | 75.68 |
Cross Lipschitz (regular training) [52] | 96.78 | 92.54 | 91.42 | 80.10 | - | - |
Cross Lipschitz (fine-tuning) [52] | 98.77 | 92.41 | 93.50 | 79.39 | - | - |