Table 5 F-score 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 | 92.66 | 94.17 | 82.93 | 96.18 | 76.49 |
GCCS (fine-tuning) | 99.64 | 93.80 | 95.28 | 81.46 | 97.05 | 77.72 |
No Defense (cross-entropy loss) | 99.35 | 91.88 | 93.7 | 78.59 | 95.77 | 76.55 |
Jacobian Reg. (regular training) [53] | 98.98 | 91.73 | 93.68 | 69.32 | - | - |
Jacobian Reg. (fine-tuning) [53] | 98.51 | 92.41 | 93.24 | 82.2 | - | - |
Input Gradient Reg. (regular training) [44] | 97.96 | 88.51 | 93.26 | 78.70 | 96.58 | 76.24 |
Input Gradient Reg. (fine-tuning) [44] | 99.08 | 92.38 | 92.62 | 76.39 | 96.98 | 75.59 |
Cross Lipschitz (regular training) [52] | 96.64 | 92.52 | 90.55 | 80.15 | - | - |
Cross Lipschitz (fine-tuning) [52] | 98.75 | 92.39 | 92.97 | 79.22 | - | - |