config.yaml

# config for shadow
CFG:
  dataset_name: CIFAR10 # CIFAR10 vs CIFAR100 selection
  model_architecture: resnet18 # names available on torchvision.models https://pytorch.org/vision/stable/models.html#classification
  topk_num_accessible_probs: 10 # topk match with accessible classes logits/probability classes number from the target model. usually top 5 for APIs
  # "We set the learning rate to 0.001, the learning rate decay to 1e − 07, and the maximum epochs of training to 100."
  num_epochs: 100 # number of shadow model train epochs
  learning_rate: 0.001
  learning_rate_decay: 0.0000001 # NOT IMPLEMENTED ON THE REPO
  weight_decay: 0.00001 # default lr: wd ratio is 0.1(https://github.com/clovaai/AdamP), but using 0.01 for small dataset
  num_shadow_models: 128 # We set the number of shadow models to 100 for the CIFAR datasets
  # We vary the size of the training set for the CIFAR datasets, to measure the difference in the attack accuracy.
  # For the CIFAR-10 dataset, we choose 2,500; 5,000; 10,000; and 15,000.
  # For the CIFAR-100 dataset, we choose 4,600; 10,520; 19,920; and 29,540.
  shadow_train_size: 2500 # number of datasets to divide CIFAR train dataset for shadow model training
  seed: 42
  val_acc_goal: -1 # shadow model's goal accuracy working as early stop. -1 for not using early stop
  early_stop_patience: 10 # 10 epochs of patience for earlystop
  input_resolution: 32 # 32 x 32 is cifar10 and cifar100 original image resolution
  train_batch_size: 256
  val_batch_size: 512
  logging_steps: 1000
  save_path: ./ckpt
  target_train_size: 7500 # fraction to divide CIFAR test dataset for target model training
  bool_pretrained: false
# config for attack model

CFG_ATTACK:
  target_model_path: ./ckpt/target_loss_ 2.4_acc5_93.68000030517578.ckpt
  attack_dset_path: ./attack/ResNet_pretrained_False_num_shadow_128_CIFAR10.csv
  attack_model_path: ./attack/CatBoostClassifier_0.8371392405063292
  test_size: 0.2
  train_epoch: 200
  learning_rate: 0.25
  n_estimators: 100
  roc_curve_path: ./assets/roc_cifar10.png