Shokri et al. attack

See the full runnable example. Read the details of the attack in the paper.

Let target_model_fn() return the target model architecture as a scikit-like classifier. The attack is white-box, meaning the attacker is assumed to know the architecture. Let NUM_CLASSES be the number of classes of the classification problem.

First, the attacker needs to train several shadow models —that mimick the target model— on different datasets sampled from the original data distribution. The following code snippet initializes a shadow model bundle, and runs the training of the shadows. For each shadow model, 2 * SHADOW_DATASET_SIZE examples are sampled without replacement from the full attacker’s dataset. Half of them will be used for control, and the other half for training of the shadow model.

from mia.estimators import ShadowModelBundle

smb = ShadowModelBundle(
X_shadow, y_shadow = smb.fit_transform(attacker_X_train, attacker_y_train)

fit_transform returns attack data X_shadow, y_shadow. Each row in X_shadow is a concatenated vector consisting of the prediction vector of a shadow model for an example from the original dataset, and the example’s class (one-hot encoded). Its shape is hence (2 * SHADOW_DATASET_SIZE, 2 * NUM_CLASSES). Each label in y_shadow is zero if a corresponding example was “out” of the training dataset of the shadow model (control), or one, if it was “in” the training.

mia provides a class to train a bundle of attack models, one model per class. attack_model_fn() is supposed to return a scikit-like classifier that takes a vector of model predictions (NUM_CLASSES, ), and returns whether an example with these predictions was in the training, or out.

from mia.estimators import AttackModelBundle

amb = AttackModelBundle(attack_model_fn, num_classes=NUM_CLASSES), y_shadow)

In place of the AttackModelBundle one can use any binary classifier that takes (2 * NUM_CLASSES, )-shape examples (as explained above, the first half of an input is the prediction vector from a model, the second half is the true class of a corresponding example).

To evaluate the attack, one must encode the data in the above-mentioned format. Let target_model be the target model, data_in the data (tuple X, y) that was used in the training of the target model, and data_out the data that was not used in the training.

from mia.estimators import prepare_attack_data

attack_test_data, real_membership_labels = prepare_attack_data(
    target_model, data_in, data_out

attack_guesses = amb.predict(attack_test_data)
attack_accuracy = np.mean(attack_guesses == real_membership_labels)