art.defences.transformer.evasion

Module implementing transformer-based defences against evasion attacks.

Defensive Distillation

class art.defences.transformer.evasion.DefensiveDistillation(classifier: CLASSIFIER_TYPE, batch_size: int = 128, nb_epochs: int = 10)

Implement the defensive distillation mechanism.

Paper link: https://arxiv.org/abs/1511.04508

__call__(x: ndarray, transformed_classifier: CLASSIFIER_TYPE) → CLASSIFIER_TYPE

Perform the defensive distillation defence mechanism and return a robuster classifier.

Parameters:

• x (ndarray) – Dataset for training the transformed classifier.

• transformed_classifier – A classifier to be transformed for increased robustness. Note that, the objective loss function used for fitting inside the input transformed_classifier must support soft labels, i.e. probability labels.

Returns:

The transformed classifier.

__init__(classifier: CLASSIFIER_TYPE, batch_size: int = 128, nb_epochs: int = 10) → None

Create an instance of the defensive distillation defence.

Parameters:

• classifier – A trained classifier.

• batch_size (int) – Size of batches.

• nb_epochs (int) – Number of epochs to use for training.

fit(x: ndarray, y: ndarray | None = None, **kwargs) → None

No parameters to learn for this method; do nothing.