Top3 solution code for AAAI2022 安全AI挑战者计划第八期：Data-Centric Robust Learning on ML Models.

Solution Dataset

• 5k images: Original cifar10 training set.
• 5k images: Original cifar10 training set with Gaussian noise.
• 10k images: Adversarial examples of the cifar10 training set attacked by AdvDrop. (5k for preactresnet18 and 5k for wideresnet. Similar to the following.)
• 10k images: Adversarial examples of the cifar10 training set attacked by DeepFool.
• 10k images: Adversarial examples of the cifar10 training set attacked by PGD.
• 10k images: Adversarial examples of the cifar10 training set attacked by AutoAttack.

Total: 50k images.

Generate the Solution Dataset

0. Environment

pip install -r requirements.txt

1. Train the two model using the original cifar10 training set.

• Change code at line 5 in gen_dataset.py: train=True.
• Get the original cifar10 training set data.npy and label.npy:

  python gen_dataset.py
  

• Change code at lines 14 and 29 in config.py: 'batch_size': 64
• Train the two model, get the trained model preactresnet18.pth.tar and wideresnet.pth.tar:

  python train.py
  

2. Split the original cifar10 training set into 10 parts.

python split.py

The training set will be split into 10 parts: data_1.npy, label_1.npy, data_2.npy, label_2.npy, ... , data_10.npy, label_10.npy.

Each part has 5k images.

3. Get each part of the Solution Dataset

3.1 Gaussian noise for data_2.npy

cd naive
python naive.py

The data will be saved in data_naive.npy and label_naive.npy

3.2 AdvDrop for data_3.npy and data_4.npy

• Change codes at lines 27-30 in 'AdvDrop-main/infod_sample.py'

  data_npy = '../data_3.npy'
  label_npy = '../label_3.npy'
  arch = 'wideresnet'
  

• Run and get solution dataset part AdvDrop-main/advdrop_train_wideresnet.npy

  cd AdvDrop-main
  python infod_sample.py
  

• Change codes at lines 27-30 in 'AdvDrop-main/infod_sample.py'

  data_npy = '../data_4.npy'
  label_npy = '../label_4.npy'
  arch = 'preactresnet18'
  

• Run and get solution dataset part AdvDrop-main/advdrop_train_preactresnet18.npy

  cd AdvDrop-main
  python infod_sample.py
  

3.3 DeepFool for data_5.npy and data_6.npy

• Change codes at lines 24-27 in 'DeepFool/Python/test_deepfool.py'

  data_npy = '../../data_5.npy'
  label_npy = '../../label_5.npy'
  arch = 'wideresnet'
  

• Run and get solution dataset part DeepFool/Python/deepfool_train_wideresnet.npy

  cd DeepFool/Python
  python test_deepfool.py
  

• Change codes at lines 24-27 in 'DeepFool/Python/test_deepfool.py'

  data_npy = '../../data_6.npy'
  label_npy = '../../label_6.npy'
  arch = 'preactresnet18'
  

• Run and get solution dataset part DeepFool/Python/deepfool_train_preactresnet18.npy

  cd DeepFool/Python
  python test_deepfool.py
  

3.4 PGD for data_7.npy and data_8.npy

• Change codes at lines 24-27 in 'pgd/pgd.py'

  data_npy = '../data_7.npy'
  label_npy = '../label_7.npy'
  arch = 'wideresnet'
  

• Run and get solution dataset part pgd/pgd_train_wideresnet.npy

  cd pgd
  python pgd.py
  

• Change codes at lines 24-27 in 'pgd/pgd.py'

  data_npy = '../data_8.npy'
  label_npy = '../label_8.npy'
  arch = 'preactresnet18'
  

• Run and get solution dataset part pgd/pgd_train_preactresnet18.npy

  cd pgd
  python pgd.py
  

3.5 AutoAttack for data_9.npy and data_10.npy

• Change codes at lines 24-27 in 'aa/aa.py'

  data_npy = '../data_10.npy'
  label_npy = '../label_10.npy'
  arch = 'wideresnet'
  

• Run and get solution dataset part aa/aa_train_wideresnet.npy

  cd aa
  python aa.py
  

• Change codes at lines 24-27 in 'aa/aa.py'

  data_npy = '../data_9.npy'
  label_npy = '../label_9.npy'
  arch = 'preactresnet18'
  

• Run and get solution dataset part aa/aa_train_preactresnet18.npy

  cd aa
  python aa.py
  

4. Combined the Solution Dataset

python mix.py

The solution dataset will be saved and overwrite the original dataset file data.npy and label.npy.

Then retrain the two model.

Change code at lines 14 and 29 in config.py: 'batch_size': 48

python train.py

Acknowledgments

• Advdrop: https://github.com/RjDuan/AdvDrop
• DeepFool: https://github.com/LTS4/DeepFool
• torchattacks: https://adversarial-attacks-pytorch.readthedocs.io/en/latest/