Skip to main content

Table 1 Training parameters

From: Secure machine learning against adversarial samples at test time

Parameter MNIST
αNormalNoise 0.44
αBIM 0.44
αCW 0.54
αDF 0.54
τNormalNoise 0.1
τBIM 0.1
τCW 0.0265
τDF 0.0265
β \(\frac {\alpha }{2}+\frac {1-\alpha }{11}\)
γ β
ρ0 0.1
Batch size 60
kmax 3000
Steps 100