Skip to main content

Table 3 Adversarial robustness measures using no regularization, dropout, and flipover respectively

From: Flipover outperforms dropout in deep learning

Model

Method

ACC Org (%)

ACC Att (%)

Para comb

Small CNN [17]

Vanilla

98.53

5.50

/

Dropout

92.43

34.98

(0.7, 0.9)

Flipover

98.00

49.79

(0.3, 0.0)

ResNet18 [4]

Vanilla

93.63

23.00

/

Dropout

93.42

23.14

(0.0, 0.3)

Flipover

92.37

40.99

(0.3, 0.0)

  1. Note: ACC Org stands for the accuracy on the original test set; ACC Att for the accuracy under adversarial attack; Para Comb for the optimal combination of training and test rates