"""
.. module:: CFoolboxPGD
:synopsis: Performs Foolbox PGD attack.
.. moduleauthor:: Luca Demetrio <luca.demetrio@dibris.unige.it>
.. moduleauthor:: Maura Pintor <maura.pintor@unica.it>
.. moduleauthor:: Giovanni Manca <g.manca72@studenti.unica.it>
"""
from foolbox.attacks.projected_gradient_descent import L1ProjectedGradientDescentAttack, \
L2ProjectedGradientDescentAttack, LinfProjectedGradientDescentAttack
from secml.adv.attacks.evasion.foolbox.c_attack_evasion_foolbox import CAttackEvasionFoolbox
from secml.adv.attacks.evasion.foolbox.losses.ce_loss import CELoss
from secml.adv.attacks.evasion.foolbox.secml_autograd import as_tensor
DISTANCES = ['l1', 'l2', 'linf']
[docs]class CFoolboxPGD(CELoss, CAttackEvasionFoolbox):
"""
Projected Gradient Descent Attack [#Madry17]_.
Credits: https://github.com/bethgelab/foolbox/blob/master/foolbox/attacks/projected_gradient_descent.py
Parameters
----------
classifier : CClassifier
Trained secml classifier.
y_target : int or None, optional
If None an indiscriminate attack will be performed, else a
targeted attack to have the samples misclassified as
belonging to the y_target class.
lb : float or None, optional
Lower bound of the model's input space.
ub : float or None, optional
Upper bound of the model's input space.
epsilons : float or None, optional
The maximum size of the perturbations, required for the
fixed epsilon foolbox attacks.
distance : str, optional
Norm of the attack. One of 'l1', 'l2', 'linf'.
rel_stepsize : float, optional
Stepsize relative to epsilon.
abs_stepsize : float, optional
If given, it takes precedence over rel_stepsize.
steps : int, optional
Number of update steps to perform.
random_start : bool, optional
Whether the perturbation is initialized randomly or starts at zero.
References
----------
.. [#Madry17] Madry, A., Makelov, A., Schmidt, L., Tsipras, D.,
& Vladu, A. "Towards deep learning models resistant to
adversarial attacks",
https://arxiv.org/abs/1706.06083
"""
__class_type = 'e-foolbox-pgd'
def __init__(self, classifier, y_target=None, lb=0.0, ub=1.0,
epsilons=0.2, distance='l2',
rel_stepsize=0.025, abs_stepsize=None, steps=50,
random_start=True):
if distance == 'l1':
attack = L1ProjectedGradientDescentAttack
elif distance == 'l2':
attack = L2ProjectedGradientDescentAttack
elif distance == 'linf':
attack = LinfProjectedGradientDescentAttack
else:
raise ValueError('Distance {} is not supported for this attack. Only {} are supported'.format(
distance, DISTANCES
))
super(CFoolboxPGD, self).__init__(classifier, y_target,
lb=lb, ub=ub,
fb_attack_class=attack,
epsilons=epsilons,
rel_stepsize=rel_stepsize,
abs_stepsize=abs_stepsize,
steps=steps,
random_start=random_start)
self._x0 = None
self._y0 = None
self.distance = distance
def _run(self, x, y, x_init=None):
self._x0 = as_tensor(x)
self._y0 = as_tensor(y)
out, _ = super(CFoolboxPGD, self)._run(x, y, x_init)
self._f_seq = self.objective_function(self.x_seq)
f_opt = self.objective_function(out)
return out, f_opt
[docs]class CFoolboxPGDL1(CFoolboxPGD):
__class_type = 'e-foolbox-pgd-l1'
def __init__(self, classifier, y_target=None, lb=0.0, ub=1.0, epsilons=0.2,
rel_stepsize=0.025, abs_stepsize=None, steps=50, random_start=True):
super(CFoolboxPGDL1, self).__init__(classifier, y_target,
lb=lb, ub=ub,
distance='l1',
epsilons=epsilons,
rel_stepsize=rel_stepsize,
abs_stepsize=abs_stepsize,
steps=steps,
random_start=random_start)
[docs]class CFoolboxPGDL2(CFoolboxPGD):
__class_type = 'e-foolbox-pgd-l2'
def __init__(self, classifier, y_target=None, lb=0.0, ub=1.0, epsilons=0.2,
rel_stepsize=0.025, abs_stepsize=None, steps=50, random_start=True):
super(CFoolboxPGDL2, self).__init__(classifier, y_target,
lb=lb, ub=ub,
distance='l2',
epsilons=epsilons,
rel_stepsize=rel_stepsize,
abs_stepsize=abs_stepsize,
steps=steps,
random_start=random_start)
[docs]class CFoolboxPGDLinf(CFoolboxPGD):
__class_type = 'e-foolbox-pgd-linf'
def __init__(self, classifier, y_target=None, lb=0.0, ub=1.0, epsilons=0.2,
rel_stepsize=0.025, abs_stepsize=None, steps=50, random_start=True):
super(CFoolboxPGDLinf, self).__init__(classifier, y_target,
lb=lb, ub=ub,
distance='linf',
epsilons=epsilons,
rel_stepsize=rel_stepsize,
abs_stepsize=abs_stepsize,
steps=steps,
random_start=random_start)