From: Deep neural rejection against adversarial examples
Parameter
MNIST model
CIFAR model
Learning rate
0.1
0.01
Momentum
0.9
Dropout
0.5
(see Table 1)
Batch size
128
100
Epochs
50
75