Table 8 Training time for regular vs adversarial training over the MNIST dataset (1800 epochs). The numbers in brackets indicate the training time percentage increase caused by the considered defense method as compared to the corresponding plain cross-entropy training case (No Defense)
From: Gaussian class-conditional simplex loss for accurate, adversarially robust deep classifier training
Applied defense method | Training time (minutes) |
|---|---|
No Defense (cross-entropy loss) | 312 |
No Defense (adversarial training) | 687 |
GCCS (regular training) | 318 (+1.92%) |
GCCS (adversarial training) | 694 (+1.02%) |
JR (regular training) [53] | 512 (+64.10%) |
JR (adversarial training) [53] | 1073 (+56.18%) |
IGR (regular training) [44] | 617 (+97.75%) |
IGR (adversarial training) [44] | 1297 (+88.79%) |
CLR (regular training) [52] | 687 (+120.2%) |
CLR (adversarial training) [52] | 1456 (+111.9%) |