search

ZiangYan / deepdefense.pytorch

Implementation of our NeurIPS 2018 paper: Deep Defense: Training DNNs with Improved Adversarial Robustness
39 stars 7 forks source link

value of Epsilon_ref in table 1 #4

Closed liangbright closed 3 years ago

liangbright commented 5 years ago

Hi Ziang Yan,

What is the value of Epsilon_ref in table 1 ?

Best, Liang

ZiangYan commented 5 years ago

Hi,

Please see the table below for MNIST and CIFAR-10 epsilon ref values.

Dataset Arch Epsilon ref
MNIST MLP 35.16
MNIST LeNet 53.98
CIFAR-10 ConvNet 2.75
CIFAR-10 NIN 3.75

Please also note the evaluation protocol differs from the current common practice in two aspects:

  1. The epsilon values are calculated on the processed images instead of the raw image (i.e., in [0, 1] or [0, 255] range) and the pre-processing procedure are somewhat non-standard (see #3 for more details).
  2. The perturbed images are not clipped back to the valid image domain.