Skip to main content

Table 3 Ratio \(r_{AB}\) of inter-class to intra-class distance 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)

18.50 ± 1.45

7.64 ± 5.42

5.37 ± 0.65

2.89 ± 1.01

9.21 ± 3.52

2.34 ± 0.87

GCCS (fine-tuning)

18.11 ± 1.59

8.29 ± 5.69

7.33 ± 0.85

2.91 ± 1.22

9.57 ± 3.16

2.16 ± 0.51

No Defense (cross-entropy loss)

3.12 ± 0.71

3.13 ± 1.24

1.94 ± 0.21

1.71 ± 0.31

2.71 ± 0.48

1.62 ± 0.26

Jacobian Reg. (regular training) [53]

3.35 ± 1.01

3.64 ± 1.22

1.94 ± 0.21

2.03 ± 0.68

-

-

Jacobian Reg. (fine-tuning)[53]

4.09 ± 0.81

3.71 ± 1.24

2.43 ± 0.26

2.35 ± 0.63

-

-

Input Gradient Reg. (regular training) [44]

3.70 ± 0.92

2.71 ± 0.85

2.12 ± 0.24

1.57 ± 0.24

2.70 ± 0.55

1.63 ± 0.27

Input Gradient Reg. (fine-tuning) [44]

3.65 ± 0.97

3.18 ± 1.03

2.11 ± 0.27

1.68 ± 0.31

2.97 ± 0.59

1.60 ± 0.25

Cross Lipschitz (regular training) [52]

4.43 ± 1.07

4.44 ± 3.24

2.40 ± 0.29

1.91 ± 0.41

-

-

Cross Lipschitz (fine-tuning) [52]

6.72 ± 2.23

4.42 ± 3.09

2.62 ± 0.28

1.85 ± 0.31

-

-