SafeModel Examples#

This section demonstrates how to use the SafeModel package for ante-hoc privacy assessment.

SafeModel Usage#

Example showing how to use SafeModel wrappers for privacy-aware machine learning:

SafeModel Usage Example#
 1"""Example showing how to integrate attacks into safemodel classes."""
 2
 3import logging
 4
 5import numpy as np
 6from sklearn.datasets import fetch_openml
 7from sklearn.model_selection import train_test_split
 8from sklearn.preprocessing import LabelEncoder, OneHotEncoder
 9
10from sacroml.attacks.target import Target
11from sacroml.safemodel.classifiers import SafeDecisionTreeClassifier
12
13output_dir = "outputs_safemodel"
14
15if __name__ == "__main__":
16    logging.info("Loading dataset")
17    nursery_data = fetch_openml(data_id=26, as_frame=True)
18    X = np.asarray(nursery_data.data, dtype=str)
19    y = np.asarray(nursery_data.target, dtype=str)
20
21    n_features = np.shape(X)[1]
22    indices = [
23        [0, 1, 2],  # parents
24        [3, 4, 5, 6, 7],  # has_nurs
25        [8, 9, 10, 11],  # form
26        [12, 13, 14, 15],  # children
27        [16, 17, 18],  # housing
28        [19, 20],  # finance
29        [21, 22, 23],  # social
30        [24, 25, 26],  # health
31    ]
32
33    logging.info("Splitting data into training and test sets")
34    X_train_orig, X_test_orig, y_train_orig, y_test_orig = train_test_split(
35        X, y, test_size=0.5, stratify=y, shuffle=True
36    )
37
38    logging.info("Preprocessing dataset")
39    label_enc = LabelEncoder()
40    feature_enc = OneHotEncoder()
41    X_train = feature_enc.fit_transform(X_train_orig).toarray()
42    y_train = label_enc.fit_transform(y_train_orig)
43    X_test = feature_enc.transform(X_test_orig).toarray()
44    y_test = label_enc.transform(y_test_orig)
45
46    logging.info("Defining the (safe) model")
47    model = SafeDecisionTreeClassifier(random_state=1)
48
49    logging.info("Training the model")
50    model.fit(X_train, y_train)
51    acc_train = model.score(X_train, y_train)
52    acc_test = model.score(X_test, y_test)
53    logging.info("Base model train accuracy: %.4f", acc_train)
54    logging.info("Base model test accuracy: %.4f", acc_test)
55
56    logging.info("Performing a preliminary check")
57    msg, disclosive = model.preliminary_check()
58
59    logging.info("Wrapping the model and data in a Target object")
60    target = Target(
61        model=model,
62        dataset_name="nursery",
63        X_train=X_train,
64        y_train=y_train,
65        X_test=X_test,
66        y_test=y_test,
67        X_train_orig=X_train_orig,
68        y_train_orig=y_train_orig,
69        X_test_orig=X_test_orig,
70        y_test_orig=y_test_orig,
71    )
72    for i in range(n_features):
73        target.add_feature(nursery_data.feature_names[i], indices[i], "onehot")
74
75    logging.info("Dataset: %s", target.dataset_name)
76    logging.info("Features: %s", target.features)
77    logging.info("X_train shape: %s", str(target.X_train.shape))
78    logging.info("y_train shape: %s", str(target.y_train.shape))
79    logging.info("X_test shape: %s", str(target.X_test.shape))
80    logging.info("y_test shape: %s", str(target.y_test.shape))
81
82    logging.info("Performing disclosure checks")
83    model.request_release(path=output_dir, ext="pkl", target=target)
84
85    logging.info("Please see the files generated in: %s", output_dir)

The SafeModel package provides wrappers around common machine learning models that automatically assess privacy risks before training. This helps researchers identify potentially problematic model configurations before investing time in training.

On this page

This Page