def testInitInvalidInput(self):
''' Handles unregistered or missing ATTACK.INIT parameter.'''
config = copy.deepcopy(self._config)
config.ATTACK.INIT.TYPE = 'ERROR'
with self.assertRaises(KeyError):
attack.IFGSM(config.ADV_MODEL, config.ATTACK)
del config.ATTACK.INIT.TYPE
with self.assertRaises(AttributeError):
attack.IFGSM(config.ADV_MODEL, config.ATTACK)
del config
def testEpsilon(self):
config = copy.deepcopy(self._config)
eps_type = {
'RANDOM': attack.RandomEpsilon,
'FIXED': attack.FixedEpsilon
}
for (eps, cls) in eps_type.items():
with self.subTest(eps=eps):
config.ATTACK.EPSILON.TYPE = eps
c = attack.IFGSM(config.ADV_MODEL, config.ATTACK)
self.assertIsInstance(c.epsManager, cls)
del config
def testInitValidInput(self):
''' Ensures correct Init subclass is called based on parameter configuration'''
config = copy.deepcopy(self._config)
init_type = {
'RAW': attack.ZeroShift,
'SHIFT': attack.RandomShift,
}
for (init, cls) in init_type.items():
with self.subTest(type=init):
config.ATTACK.INIT.TYPE = init
c = attack.IFGSM(config.ADV_MODEL, config.ATTACK)
self.assertIsInstance(c.initManager, cls)
del config
def testAlpha(self):
''' Ensures correct Alpha subclass is called based on parameter configuration.'''
config = copy.deepcopy(self._config)
alpha_type = {
'DIVISOR': attack.DivisorAlpha,
}
for (alpha, cls) in alpha_type.items():
with self.subTest(alpha=alpha):
config.ATTACK.ALPHA.TYPE = alpha
c = attack.IFGSM(config.ADV_MODEL, config.ATTACK)
self.assertIsInstance(c.alphaManager, cls)
del config
def testRun(self):
''' Test adversarial example generation. Constraint to a batch_size of 1.'''
# 1. Hardcoded adversarial example (torch tensor format)
x = torch.load('raw_image_ts.pt').unsqueeze(0)
y = torch.cuda.LongTensor([3])
iter2adv = {
1: torch.load('adv_image_ts.pt').unsqueeze(0),
4: torch.load('adv4_image_ts.pt').unsqueeze(0),
10: torch.load('adv10_image_ts.pt').unsqueeze(0)
}
for (n_iter, adv) in iter2adv.items():
with self.subTest(n_iter=n_iter):
config = copy.deepcopy(self._config)
config.ATTACK.MAX_ITER = n_iter
config.ATTACK.ALPHA.DIVISOR = n_iter
c = attack.IFGSM(config.ADV_MODEL, config.ATTACK)
# Tensor Equality
self.assertTrue(torch.all(adv.eq(c.run(x, y))))