Skip to main content

Table 6 Robustness to DeepFool attack obtained through regular vs adversarial training over different benchmark datasets with different competing techniques, measured with \(\rho\) [32]

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)

3.79

4.98

1.85

2.53

0.46

0.0034

GCCS (adversarial training)

3.93

4.91

3.05

2.42

0.94

0.0051

No Defense (cross-entropy loss)

1.83

0.55

1.73

1.43

0.27

0.0011

No Defense (adversarial training)

1.10

0.45

0.52

0.48

0.06

0.0013

JR (regular training) [53]

0.55

0.31

1.73

0.32

-

-

JR (adversarial training) [53]

1.60

0.39

0.53

0.17

-

-

IGR (regular training) [44]

0.48

0.37

1.54

0.83

0.05

0.0013

IGR (adversarial training) [44]

1.79

0.46

0.58

0.46

0.06

0.0017

CLR (regular training) [52]

0.31

0.29

0.33

0.26

-

-

CLR (adversarial training) [52]

0.85

0.39

0.36

0.26

-

-