def __new__(cls, *args, **kwargs):
cls.name = "parametric_defense"
grid_params = []
grid_params.append({
'model': [
'logistic_regression',
'adv_logistic_regression_50',
'advPruning_logistic_regression_50',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['pgd'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'mlp',
'adv_mlp_50',
'advPruning_mlp_50',
],
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': ['pgd'],
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf-robustness"
attacks = ['RBA_Approx_RF_100']
invert_ds_eps = {}
for k, v in ds_eps.items():
invert_ds_eps.setdefault(v, []).append(
auto_var.get_var_shown_name("dataset", k))
grid_params = []
for ds in tree_datasets:
v, k = [ds], ds_eps[auto_var.get_var_shown_name("dataset", ds)]
models = [
'random_forest_100_d5',
f'adv_rf_100_{k}_d5',
f'robust_rf_100_{k}_d5',
f'advPruning_rf_100_{k}_d5',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': v,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "dt-robustness"
attacks = ['pgd']
invert_ds_eps = {}
for k, v in ds_eps.items():
invert_ds_eps.setdefault(v, []).append(
auto_var.get_var_shown_name("dataset", k))
grid_params = []
for ds in tree_datasets:
v, k = [ds], ds_eps[auto_var.get_var_shown_name("dataset", ds)]
models = [
'logistic_regression',
f'adv_logistic_regression_{k}',
f'advPruning_logistic_regression_{k}',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': v,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "compare_nns"
grid_params = []
for i in [1, 3, 5, 7]:
grid_params.append({
'model': ['knn%d' % i],
'ord': [ATTACK_NORM],
'dataset': datasets,
'attack': ['RBA_Approx_KNN_k%d_50' % i],
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "nn_k3_robustness_figs"
nn_k3_models = [
'knn3',
'advPruning_nn_k3_30',
]
grid_params = {
'model': nn_k3_models,
'attack': ['RBA_Approx_KNN_k3_50'],
'dataset': datasets,
'ord': [ATTACK_NORM],
'random_seed': random_seed,
}
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "compare_attacks"
grid_params = []
grid_params.append({
'model': ['knn1'],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': [
'direct_k1',
'blackbox',
'kernelsub_c1000_pgd',
'RBA_Exact_KNN_k1',
'RBA_Approx_KNN_k1_50',
],
'random_seed':
random_seed,
})
grid_params.append({
'model': ['knn3'],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': [
'direct_k3', 'blackbox', 'kernelsub_c1000_pgd',
'RBA_Approx_KNN_k3_50'
],
'random_seed':
random_seed,
})
grid_params.append({
'model': ['decision_tree_d5'],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['dt_papernots', 'blackbox', 'RBA_Exact_DT'],
'random_seed':
random_seed,
})
grid_params.append({
'model': ['random_forest_100_d5'],
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': ['blackbox', 'RBA_Approx_RF_100'],
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf_robustness_figs"
rf_models = [
'random_forest_100_d5',
'robust_rf_100_30_d5',
'advPruning_rf_100_30_d5',
]
grid_params = {
'model': rf_models,
'attack': ['RBA_Approx_RF_100'],
'dataset': tree_datasets,
'ord': [ATTACK_NORM],
'random_seed': random_seed,
}
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
dt_models = [
'decision_tree_d5',
'robust_decision_tree_d5_30',
'advPruning_decision_tree_d5_30',
]
cls.name = "dt_robustness_figs"
grid_params = {
'model': dt_models,
'attack': ['RBA_Exact_DT'],
'dataset': tree_datasets,
'ord': [ATTACK_NORM],
'random_seed': random_seed,
}
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "dt-robustness"
attacks = ['pgd']
grid_params = []
for ds in tree_datasets:
v, k = [ds], ds_eps[auto_var.get_var_shown_name("dataset", ds)]
models = ['mlp', f'adv_mlp_{k}', f'advPruning_mlp_{k}']
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': v,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "3nn_robustness"
grid_params = []
grid_params.append({
'model': [
'knn3',
'advPruning_nn_k3_10',
'advPruning_nn_k3_30',
'advPruning_nn_k3_50',
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Approx_KNN_k3_50'], #, 'blackbox'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "fullds"
grid_params = []
grid_params.append({
'model': [
'random_forest_500_d10',
'approxAP_rf_500_20_d10',
'xgb_xgb_models/prune_fullmnist_pca100_20_rf_linf.0200.model',
"knn1",
"approxAP_nn_k1_20",
"knn3",
"approxAP_nn_k3_20",
'xgb_xgb_models/prune_fullmnist_pca100_20_linf.unrob.0200.model',
'xgb_xgb_models/fullmnist_pca100.unrob.0200.model',
'xgb_xgb_models/fullmnist_pca100.0200.model',
],
'ord': [ATTACK_NORM],
'dataset': ['fullmnist_pca100'],
'attack': ['blackbox'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'random_forest_500_d10',
'approxAP_rf_500_20_d10',
'xgb_xgb_models/prune_fullfashion_pca100_20_rf_linf.0200.model',
"knn1",
"approxAP_nn_k1_20",
"knn3",
"approxAP_nn_k3_20",
'xgb_xgb_models/prune_fullfashion_pca100_20_linf.unrob.0200.model',
'xgb_xgb_models/fullfashion_pca100.unrob.0200.model',
'xgb_xgb_models/fullfashion_pca100.0200.model',
],
'ord': [ATTACK_NORM],
'dataset': ['fullfashion_pca100'],
'attack': ['blackbox'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "dt-robustness"
models = [
'decision_tree_d5',
'advPruning_decision_tree_d5_10',
'advPruning_decision_tree_d5_30',
'advPruning_decision_tree_d5_50',
]
attacks = ['RBA_Exact_DT']
grid_params = []
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "fullds"
grid_params = []
grid_params.append({
'model': [
'faisslshknn_3_200',
'approxAP_faisslshknn_3_200_500',
'random_forest_100_d10',
'approxAP_rf_100_500_d10',
'mlp',
'adv_mlp_500',
],
'ord': [ATTACK_NORM],
'dataset': ['fullmnist', 'fullfashion'],
'attack': ['blackbox'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "mlp-at-robustness"
models = [
'mlp',
'adv_mlp_25',
'adv_mlp_50',
'adv_mlp_75',
]
attacks = ['pgd']
grid_params = []
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "fullds_l2"
grid_params = []
grid_params.append({
'model': [
'random_forest_300_d10',
'approxAP_rf_300_500_d10',
"knn1",
"approxAP_nn_k1_500",
"knn3",
"approxAP_nn_k3_500",
],
'ord': [ATTACK_NORM],
'dataset': ['fullmnist', 'fullfashion'],
'attack': ['blackbox'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "lr-ap-robustness"
models = [
'logistic_regression',
'advPruning_logistic_regression_25',
'advPruning_logistic_regression_50',
'advPruning_logistic_regression_75',
]
attacks = ['pgd']
grid_params = []
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "1nn_robustness"
grid_params = []
grid_params.append({
'model': [
'knn1',
'advPruning_nn_k1_25',
'advPruning_nn_k1_50',
'advPruning_nn_k1_75',
#'adv_nn_k1_10', 'adv_nn_k1_30', 'adv_nn_k1_50',
#'robustv2_nn_k1_10', 'robustv2_nn_k1_30', 'robustv2_nn_k1_50',
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Exact_KNN_k1'], #, 'blackbox'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "dt-robustness"
attacks = ['pgd']
grid_params = []
for ds in tree_datasets:
k = 30
models = [
'logistic_regression',
f'adv_logistic_regression_{k}',
f'advPruning_logistic_regression_{k}',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': [ds],
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf-robustness"
attacks = ['RBA_Approx_KNN_k3_50']
grid_params = []
for ds in datasets:
k = 30
models = [
'knn3',
f'adv_nn_k3_{k}',
f'robustv2_nn_k3_{k}',
f'advPruning_nn_k3_{k}',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': [ds],
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf-robustness"
attacks = ['RBA_Approx_RF_100']
grid_params = []
for ds in tree_datasets:
k = 30
models = [
'random_forest_100_d5',
f'adv_rf_100_{k}_d5',
f'robust_rf_100_{k}_d5',
f'advPruning_rf_100_{k}_d5',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': [ds],
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf-robustness_l2"
attacks = ['RBA_Approx_KNN_k3_50']
grid_params = []
for ds in datasets:
v, k = [ds], ds_eps[auto_var.get_var_shown_name("dataset", ds)]
models = [
'knn3',
f'adv_nn_k3_{k}',
f'robustv2_nn_k3_{k}',
f'advPruning_nn_k3_{k}',
]
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': v,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "rf-robustness"
models = [
'random_forest_100_d5',
#'adv_rf_100_10_d5', 'adv_rf_100_30_d5', 'adv_rf_100_50_d5',
#'robust_rf_100_10_d5', 'robust_rf_100_30_d5', 'robust_rf_100_50_d5',
'advPruning_rf_100_25_d5',
'advPruning_rf_100_50_d5',
'advPruning_rf_100_75_d5',
#'robustv2_rf_100_10_d5', 'robustv2_rf_100_30_d5', 'robustv2_rf_100_50_d5',
]
attacks = ['RBA_Approx_RF_100']
grid_params = []
grid_params.append({
'model': models,
'ord': [ATTACK_NORM],
'dataset': tree_datasets,
'attack': attacks,
'random_seed': random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "compare_defense"
def_strength = 50
grid_params = []
grid_params.append({
'model': [
'knn1', f'adv_nn_k1_{def_strength}',
f'robustv2_nn_k1_{def_strength}',
f'advPruning_nn_k1_{def_strength}'
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Exact_KNN_k1'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'knn3', f'adv_nn_k3_{def_strength}',
f'advPruning_nn_k3_{def_strength}'
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Approx_KNN_k3_50'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'decision_tree_d5',
f'adv_decision_tree_d5_{def_strength}',
f'robust_decision_tree_d5_{def_strength}',
f'advPruning_decision_tree_d5_{def_strength}',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['RBA_Exact_DT'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'random_forest_100_d5',
f'adv_rf_100_{def_strength}_d5',
f'robust_rf_100_{def_strength}_d5',
f'advPruning_rf_100_{def_strength}_d5',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['RBA_Approx_RF_100'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'logistic_regression',
f'adv_logistic_regression_{def_strength}',
f'advPruning_logistic_regression_{def_strength}',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['pgd'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'mlp',
f'adv_mlp_{def_strength}',
f'advPruning_mlp_{def_strength}',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['pgd'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
def __new__(cls, *args, **kwargs):
cls.name = "tst_scores"
grid_params = []
grid_params.append({
'model': [
#'knn1', 'advPruning_nn_k1_10', 'advPruning_nn_k1_30', 'advPruning_nn_k1_50',
'knn1',
'advPruning_nn_k1_25',
'advPruning_nn_k1_50',
'advPruning_nn_k1_75',
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Exact_KNN_k1'], #, 'blackbox'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'knn3',
'advPruning_nn_k3_25',
'advPruning_nn_k3_50',
'advPruning_nn_k3_75',
],
'ord': [ATTACK_NORM],
'dataset':
datasets,
'attack': ['RBA_Approx_KNN_k3_50'], #, 'blackbox'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'decision_tree_d5',
'advPruning_decision_tree_d5_25',
'advPruning_decision_tree_d5_50',
'advPruning_decision_tree_d5_75',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['RBA_Exact_DT'],
'random_seed':
random_seed,
})
grid_params.append({
'model': [
'random_forest_100_d5',
'advPruning_rf_100_25_d5',
'advPruning_rf_100_50_d5',
'advPruning_rf_100_75_d5',
],
'ord': [ATTACK_NORM],
'dataset':
tree_datasets,
'attack': ['RBA_Approx_RF_100'],
'random_seed':
random_seed,
})
cls.grid_params = grid_params
return RobustExperiments.__new__(cls, *args, **kwargs)
