PyTorch Implementation of Dual Manifold Adversarial Training (DMAT)

Requirements

• CUDA 10.0.130 or newer, cudnn v7.6.5 or newer.
• Python 3.6.
• PyTorch 1.3.1 or newer with GPU support.
• tqdm
• advertorch (pip install advertorch)

Dataset and trained StyleGAN

You can download the dataset and trained StyleGAN by the link.

Structure of the configs

Since we need to handle GAN, classifier training, and adversarial attack in the same place, it's important to have a structured configuration. Take classifier training as an example, in experiments/encoders/mnist.yml:

num_epochs: 20

classifier:
  name: imagenet_classifier
  path: runs/classifiers/mixed10_pgd5_pgd5_sgd_sam

optimizer:
  module: torch.optim
  name: SGD
  args:
    lr: 0.1
    momentum: 0.9
    weight_decay: 0.0001

scheduler:
  type: cyclic
  args:
    lr_epochs: !!python/tuple [0,1,6,12,15,20]
    lr_values: !!python/tuple [0,0.4,0.04,0.004,0.0004,0.0001]

image_attack:
  module: advertorch.attacks
  name: PGDAttack
  args:
    eps: 0.03137255 # 8/255 in [-1, 1]
    eps_iter: 0.00784314 # 2/255 in [-1, 1]
    nb_iter: 5
    clip_min: -1.0
    clip_max: 1.0

latent_attack:
  module: advertorch.attacks
  name: PGDAttack
  args:
    eps: 0.02
    eps_iter: 0.005
    nb_iter: 5
    clip_min: null
    clip_max: null

dataset:
  name: mixed10
  num_classes: 10
  train:
    path: /vulcanscratch/cplau/Code/Python/InvGAN-PyTorch-master/stylegan_old/dataset
    batch_size: 32
  test:
    path: /vulcanscratch/cplau/Code/Python/InvGAN-PyTorch-master/stylegan_old/dataset
    batch_size: 32

...

• num_epoch: Number of epoch
• classifier.path: the directory that the training results are saved, which include checkpoints, and images.
• classifier.name: the .py file that the generator architecture is defined. Since it is a classifier, the .py file is assumed to be under classifier.
• optimizer: Setting of optimizer
• image_attack: Setting of adversarial attack in image space
• image_attack.name: Type of attack
• image_attack.args: Include eps (attack budget), eps_iter (step size of each iteration), nb_iter (number of iteration), clip (range of the perturbed image)
• latent_attack: Setting of adversarial attack in latent space
• dataset: Setting of dataset

If we want to perform experiments to find proper hyperparameters (e.g. sweeping over different learning rates). It's better to create a new config file for each hyperparameter.

Preparing Classifiers

We train classifiers from scratch for OM_ImageNet.

> python prepare_classifiers_mix10.py --config experiments/classifiers/mixed10_adv_pgd5_pgd5_sgd_sam_clean.yml

Evaluation

After we trained the classifier, we can use various attacks to evaluate the classifier. See the scripts in script.