Safety Metrics

class perceptron.benchmarks.AdditiveNoiseMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Base class for metric that tests models against additive noise.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, epsilons=10000)

Adds uniform or Gaussian noise to the image, gradually increasing the standard deviation until the image is misclassified.

Parameters:

adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

annotation : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the Gaussian blur or number of standard deviations between 0 and 1 that should be tried.

class perceptron.benchmarks.AdditiveGaussianNoiseMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against Gaussian noise.

class perceptron.benchmarks.AdditiveUniformNoiseMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against uniform noise.

class perceptron.benchmarks.BlendedUniformNoiseMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Blends the image with a uniform noise image until it is misclassified.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, epsilons=10000, max_directions=1000)

Metric that tests models against blended uniform noise.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the blended noise or number of standard deviations between 0 and 1 that should be tried.

max_directions : int

Maximum number of random images to try.

class perceptron.benchmarks.GaussianBlurMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against Gaussian blurs.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, epsilons=10000)

Blurs the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the Gaussian blur or number of standard deviations between 0 and 1 that should be tried.

class perceptron.benchmarks.BrightnessMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against brightness variations.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, verify=False, epsilons=1000)

Change the brightness of the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input. Must be passed if a is a numpy.ndarray.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

verify : bool

If True, return verifiable bound.

epsilons : int or Iterable[float]

Either Iterable of contrast levels or number of brightness factors between 1 and 0 that should be tried. Epsilons are one minus the brightness factor. Epsilons are not used if verify = True.

class perceptron.benchmarks.ContrastReductionMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against brightness variations.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, threshold=1.0, epsilons=1000)

Reduces the contrast of the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

threshold : float

Upper bound for contrast factor

epsilons : int or Iterable[float]

Either Iterable of contrast levels or number of contrast levels between 1 and 0 that should be tried. Epsilons are one minus the contrast level.

class perceptron.benchmarks.MotionBlurMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Motion blurs the image until it is misclassified.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, motion_angle=0, epsilons=10000)

Blurs the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

motion_angle : float

Motion angle in degree between 0 and 180.

epsilons : int or Iterable[float]

Either Iterable of kernel size that should be tried.

class perceptron.benchmarks.RotationMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against rotations.

__call__(self, adv, ang_range=None, annotation=None, unpack=True, abort_early=True, verify=False, epsilons=1000)

Rotate the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

ang_range : (float, float)

Range of angles for rotation metric.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

verify : bool

if True, return verifiable bound.

epsilons : int or Iterable[float]

Either Iterable of rotation angles or number of angles between 1 and 0 that should be tried. Epsilons are one minus the contrast level. Epsilons are not used if verify = True.

class perceptron.benchmarks.SaltAndPepperNoiseMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Add salt and pepper noise.

__call__(self, adv, annotation=None, unpack=True, abort_early=True, epsilons=10000, repetitions=10)

Add salt and pepper noise.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the salt and pepper or number of standard deviations between 0 and 1 that should be tried.

repetitions : int

Specifies how often the attack will be repeated.

class perceptron.benchmarks.SpatialMetric(model=None, criterion=None, distance=<class 'perceptron.utils.distances.MeanSquaredDistance'>, threshold=None)

Metric that tests models against spatial transformations.

__call__(self, adv, annotation=None, do_rotations=True, do_translations=True, x_shift_limits=(-5, 5), y_shift_limits=(-5, 5), angular_limits=None, unpack=True, abort_early=True, verify=False, epsilons=1000)

Spatial transforms the image until it is misclassified.

Parameters:

adv : numpy.ndarray

The original, unperturbed input as a numpy.ndarray.

annotation : int

The reference label of the original input.

do_rotations : bool

If False no rotations will be applied to the image (default True).

do_translations : bool

If False no translations will be applied to the image (default True).

x_shift_limits : (int, int)

Limits for horizontal translations in pixels (default (-5, 5)).

y_shift_limits : (int, int)

Limits for vertical translations in pixels (default (-5, 5)).

angular_limits : (int, int)

Limits for rotations in degrees (default None).

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

abort_early : bool

If true, returns when got first adversarial, otherwise returns when all the iterations are finished.

verify : bool

if True, return verifiable bound

epsilons : int or Iterable[float]

Either Iterable of contrast levels or number of contrast levels between 1 and 0 that should be tried. Epsilons are one minus the contrast level.