class FazlybSDPAdaptor(RealAdaptorBase):
"""
SDP from Fazlyb et al
"""
def __init__(self, dataset, model, timeout=30):
super(FazlybSDPAdaptor, self).__init__(dataset, model)
cp.settings.SOLVE_TIME = timeout
def prepare_solver(self, in_shape):
self.prebound = IntervalFastLinBound(self.new_model, in_shape,
self.in_min, self.in_max)
def verify(self, input, label, norm_type, radius):
"""
Here we overwrite the base class verify() method
"""
in_shape = list(input.shape)
# only support Linfty norm
assert norm_type == 'inf'
if self.new_model is None:
# init at the first time
before = time()
print(f"Init model for {self.__class__.__name__}...")
self.build_new_model(input)
self.prepare_solver(in_shape)
after = time()
print(
"Init done, time",
str(datetime.timedelta(seconds=(after - before))),
)
# firstly check the clean prediction
input = self.input_preprocess(input)
xs = input.unsqueeze(0)
clean_preds = self.model(xs.cuda()).detach().cpu().numpy()
clean_pred = np.argmax(clean_preds[0])
if clean_pred != label:
return False
m_radius = radius / self.coef
input = input.contiguous().view(-1)
self.prebound.calculate_bound(input, m_radius)
bl = [
np.maximum(self.prebound.l[i], 0) if i > 0 else self.prebound.l[i]
for i in range(len(self.prebound.l))
]
bu = [
np.maximum(self.prebound.u[i], 0) if i > 0 else self.prebound.u[i]
for i in range(len(self.prebound.u))
]
pv = BaselinePointVerifierExt(self.new_model, in_shape, self.in_min,
self.in_max)
for i in range(get_num_classes(self.dataset)):
if i != label:
pv.create_cmat(input, label, i, m_radius, bl, bu)
pv.run()
if pv.prob.status not in ['unbounded', 'unbounded_inaccurate'
] and pv.prob.value > 0.:
return False
return True
def prepare_solver(self, in_shape):
self.prebound = IntervalFastLinBound(self.new_model, in_shape,
self.in_min, self.in_max)
class MILPAdaptor(RealAdaptorBase):
"""
MILP from Tjeng et al
"""
def __init__(self, dataset, model, timeout=30):
super(MILPAdaptor, self).__init__(dataset, model)
self.timeout = timeout
def prepare_solver(self, in_shape):
self.prebound = IntervalFastLinBound(self.new_model, in_shape,
self.in_min, self.in_max)
self.bound = MILPVerifier(self.new_model, in_shape, self.in_min,
self.in_max)
def verify(self, input, label, norm_type, radius):
"""
Here we overwrite the base class verify() method
"""
# only support Linfty norm
assert norm_type == 'inf'
if self.new_model is None:
# init at the first time
before = time()
print(f"Init model for {self.__class__.__name__}...")
in_shape = list(input.shape)
self.build_new_model(input)
self.prepare_solver(in_shape)
after = time()
print(
"Init done, time",
str(datetime.timedelta(seconds=(after - before))),
)
# firstly check the clean prediction
input = self.input_preprocess(input)
xs = input.unsqueeze(0)
clean_preds = self.model(xs.cuda()).detach().cpu().numpy()
clean_pred = np.argmax(clean_preds[0])
if clean_pred != label:
return False
m_radius = radius / self.coef
input = input.contiguous().view(-1)
self.prebound.calculate_bound(input, m_radius)
self.bound.construct(self.prebound.l, self.prebound.u, input, m_radius)
for i in range(get_num_classes(self.dataset)):
if i != label:
self.bound.prepare_verify(label, i)
try:
# self.bound.prob.solve(verbose=True)
# model.setParam(GRB.Param.TimeLimit, timeout)
self.bound.prob.solve(solver=cp.GUROBI,
verbose=False,
BestObjStop=-1e-6,
TimeLimit=70,
Threads=20)
except:
return False
print(self.bound.prob.status,
self.bound.prob.value,
file=sys.stderr)
if self.bound.prob.status not in [
'optimal'
] or self.bound.prob.value < 0.:
return False
return True