art.defences.postprocessor

Module implementing postprocessing defences against adversarial attacks.

Base Class Postprocessor

class art.defences.postprocessor.Postprocessor(is_fitted: bool = False, apply_fit: bool = True, apply_predict: bool = True)

Abstract base class for postprocessing defences. Postprocessing defences are not included in the loss function evaluation for loss gradients or the calculation of class gradients.

abstract __call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(is_fitted: bool = False, apply_fit: bool = True, apply_predict: bool = True) → None

Create a postprocessing object.

Optionally, set attributes.

__weakref__

list of weak references to the object (if defined)

property apply_fit: bool

Property of the defence indicating if it should be applied at training time.

Returns:

True if the defence should be applied when fitting a model, False otherwise.

property apply_predict: bool

Property of the defence indicating if it should be applied at test time.

Returns:

True if the defence should be applied at prediction time, False otherwise.

fit(preds: ndarray, **kwargs) → None

Fit the parameters of the postprocessor if it has any.

Parameters:

• preds (ndarray) – Training set to fit the postprocessor.

• kwargs – Other parameters.

property is_fitted: bool

Return the state of the postprocessing object.

Returns:

True if the postprocessing model has been fitted (if this applies).

set_params(**kwargs) → None

Take in a dictionary of parameters and apply checks before saving them as attributes.

Class Labels

class art.defences.postprocessor.ClassLabels(apply_fit: bool = False, apply_predict: bool = True)

Implementation of a postprocessor based on adding class labels to classifier output.

__call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(apply_fit: bool = False, apply_predict: bool = True) → None

Create a ClassLabels postprocessor.

Parameters:

• apply_fit (bool) – True if applied during fitting/training.

• apply_predict (bool) – True if applied during predicting.

Gaussian Noise

class art.defences.postprocessor.GaussianNoise(scale: float = 0.2, apply_fit: bool = False, apply_predict: bool = True)

Implementation of a postprocessor based on adding Gaussian noise to classifier output.

__call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(scale: float = 0.2, apply_fit: bool = False, apply_predict: bool = True) → None

Create a GaussianNoise postprocessor.

Parameters:

• scale (float) – Standard deviation of the distribution.

• apply_fit (bool) – True if applied during fitting/training.

• apply_predict (bool) – True if applied during predicting.

High Confidence

class art.defences.postprocessor.HighConfidence(cutoff: float = 0.25, apply_fit: bool = False, apply_predict: bool = True)

Implementation of a postprocessor based on selecting high confidence predictions to return as classifier output.

__call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(cutoff: float = 0.25, apply_fit: bool = False, apply_predict: bool = True) → None

Create a HighConfidence postprocessor.

Parameters:

• cutoff (float) – Minimal value for returned prediction output.

• apply_fit (bool) – True if applied during fitting/training.

• apply_predict (bool) – True if applied during predicting.

Reverse Sigmoid

class art.defences.postprocessor.ReverseSigmoid(beta: float = 1.0, gamma: float = 0.1, apply_fit: bool = False, apply_predict: bool = True)

Implementation of a postprocessor based on adding the Reverse Sigmoid perturbation to classifier output.

__call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(beta: float = 1.0, gamma: float = 0.1, apply_fit: bool = False, apply_predict: bool = True) → None

Create a ReverseSigmoid postprocessor.

Parameters:

• beta (float) – A positive magnitude parameter.

• gamma (float) – A positive dataset and model specific convergence parameter.

• apply_fit (bool) – True if applied during fitting/training.

• apply_predict (bool) – True if applied during predicting.

Rounded

class art.defences.postprocessor.Rounded(decimals: int = 3, apply_fit: bool = False, apply_predict: bool = True)

Implementation of a postprocessor based on rounding classifier output.

__call__(preds: ndarray) → ndarray

Perform model postprocessing and return postprocessed output.

Return type:

ndarray

Parameters:

preds (ndarray) – model output to be postprocessed.

Returns:

Postprocessed model output.

__init__(decimals: int = 3, apply_fit: bool = False, apply_predict: bool = True) → None

Create a Rounded postprocessor.

Parameters:

• decimals (int) – Number of decimal places after the decimal point.

• apply_fit (bool) – True if applied during fitting/training.

• apply_predict (bool) – True if applied during predicting.