def ifgs(model,
input,
target,
max_iter=10,
step_size=0.01,
train_mode=False,
mode=None,
verbose=True):
if train_mode:
model.train()
else:
model.eval()
pred = util.get_labels(model, input)
corr = pred.eq(target)
r = torch.zeros(input.size())
for _ in range(max_iter):
_, ri = fgs(model,
input,
target,
step_size,
train_mode,
mode,
verbose=verbose)
r = r + ri
input = input + ri
pred_xp = util.get_labels(model, input + r)
status = torch.zeros(input.size(0)).long()
status[corr] = 2 * pred[corr].ne(pred_xp[corr]).long() - 1
return (status, r)
def universal(model,
input,
target,
n_classes,
max_val=0.1,
train_mode=False,
max_iter=10,
step_size=0.1,
batch_size=25,
data_dir=None,
r=None,
verbose=True):
pred = util.get_labels(model, input, batch_size)
if r is None:
r = torch.zeros(input[0].size())
perm = torch.randperm(input.size(0))
for i in range(input.size(0)):
idx = perm[i]
if verbose:
print('sample %d: index %d' % (i + 1, idx))
x_adv = torch.autograd.Variable((input[idx] + r))
x_adv = x_adv.expand(1, input.size(1), input.size(2), input.size(3))
output = model.forward(x_adv)
_, pred_adv = output.max(1)
pred_adv = pred_adv.data.cpu()[0][0]
if pred[idx] == pred_adv:
succ, ri = deepfool_single(model, input[idx] + r, pred[idx],
n_classes, train_mode, max_iter,
step_size, batch_size, data_dir)
if succ:
r = (r + ri).clamp(-max_val, max_val)
x = input + r.expand_as(input)
pred_xp = util.get_labels(model, x)
status = 2 * pred_xp.ne(target).long() - 1
status[pred.ne(target)] = 0
return (status, r)
def deepfool(model,
input,
target,
n_classes,
train_mode=False,
max_iter=5,
step_size=0.1,
batch_size=25,
labels=None):
pred = util.get_labels(model, input, batch_size)
status = torch.zeros(input.size(0)).long()
r = torch.zeros(input.size())
for i in range(input.size(0)):
status[i], r[i] = deepfool_single(model, input[i], target[i],
n_classes, train_mode, max_iter,
step_size, batch_size, labels)
status = 2 * status - 1
status[pred.ne(target)] = 0
return (status, r)
def attacking_mahalanobis_all(model,
input,
target,
weight,
loss_str,
sample_mean,
inv_sample_conv,
ensemble_weight,
bound=0,
max_iter=100,
step_size=0.01,
kappa=0,
p=2,
drop_rate=0.0,
train_mode=False,
verbose=False):
is_gpu = next(model.parameters()).is_cuda
if train_mode:
model.train()
else:
model.eval()
pred = util.get_labels(model, input)
corr = pred.eq(target)
w = torch.autograd.Variable(input, requires_grad=True)
best_w = torch.Tensor(input.size())
best_w.copy_(input)
best_loss = float('inf')
optimizer = torch.optim.Adam([w], lr=step_size)
input_var = torch.autograd.Variable(input)
input_vec = input.view(input.size(0), -1)
probs = util.get_probs(model, input)
_, top2 = probs.topk(2, 1)
argmax = top2[:, 0]
top2_index = top2[:, 1]
for j in range(top2.size(0)):
if argmax[j] == target[j]:
argmax[j] = top2[j, 1]
mahalanobis_criterion = _MahalanobisEnsembleLoss().cuda()
print('attack all')
for i in range(max_iter):
if i > 0:
w.grad.data.fill_(0)
model.zero_grad()
# Norm constraints
if loss_str == 'l2':
loss = torch.pow(w - input_var, 2).sum()
elif loss_str == 'linf':
loss = torch.clamp((w - input_var).abs() - bound, min=0).sum()
else:
raise ValueError('Unsupported loss: %s' % loss_str)
recons_loss = loss.data[0]
# Adversarial loss
w_data = w.data
w_in = torch.autograd.Variable(w_data, requires_grad=True)
output, out_features = model.feature_list(w_in)
for j in range(len(out_features)):
out_features[j] = out_features[j].view(out_features[j].size(0),
out_features[j].size(1), -1)
out_features[j] = torch.mean(out_features[j], 2)
for j in range(output.size(0)):
loss += weight * torch.clamp(
output[j][target[j]] - output[j][argmax[j]] + kappa, min=0)
temp_loss = mahalanobis_criterion(out_features, sample_mean,
inv_sample_conv, ensemble_weight,
top2_index)
adv_loss = loss.data[0] - recons_loss
if is_gpu:
loss = loss.cuda()
loss.backward()
w.grad.data.add_(w_in.grad.data)
w_cpu = w.data.cpu().numpy()
input_np = input.cpu().numpy()
optimizer.step()
total_loss = loss.data.cpu()[0]
# w.data = utils.img_to_tensor(utils.transform_img(w.data), scale=False)
output_vec = w.data
preds = util.get_labels(model, output_vec)
output_vec = output_vec.view(output_vec.size(0), -1)
diff = (input_vec - output_vec).norm(2, 1).squeeze()
diff = diff.div(input_vec.norm(2, 1).squeeze())
rb = diff.mean()
sr = float(preds.ne(target).sum()) / target.size(0)
if verbose:
print('iteration %d: loss = %f, %s_loss = %f, '
'adv_loss = %f' %
(i + 1, total_loss, loss_str, recons_loss, adv_loss))
print('robustness = %f, success rate = %f' % (rb, sr))
if total_loss < best_loss:
best_loss = total_loss
best_w = w.data.clone()
pred_xp = util.get_labels(model, best_w)
status = torch.zeros(input.size(0)).long()
status[corr] = 2 * pred[corr].ne(pred_xp[corr]).long() - 1
return (status, best_w)
def cw(model,
input,
target,
weight,
loss_str,
bound=0,
tv_weight=0,
max_iter=100,
step_size=0.01,
kappa=0,
p=2,
crop_frac=1.0,
drop_rate=0.0,
train_mode=False,
verbose=False):
is_gpu = next(model.parameters()).is_cuda
if train_mode:
model.train()
else:
model.eval()
pred = util.get_labels(model, input)
corr = pred.eq(target)
w = torch.autograd.Variable(input, requires_grad=True)
best_w = torch.Tensor(input.size())
best_w.copy_(input)
best_loss = float('inf')
optimizer = torch.optim.Adam([w], lr=step_size)
input_var = torch.autograd.Variable(input)
input_vec = input.view(input.size(0), -1)
to_pil = trans.ToPILImage()
scale_up = trans.Resize((w.size(2), w.size(3)))
scale_down = trans.Resize(
(int(crop_frac * w.size(2)), int(crop_frac * w.size(3))))
to_tensor = trans.ToTensor()
probs = util.get_probs(model, input)
_, top2 = probs.topk(2, 1)
argmax = top2[:, 0]
for j in range(top2.size(0)):
if argmax[j] == target[j]:
argmax[j] = top2[j, 1]
for i in range(max_iter):
if i > 0:
w.grad.data.fill_(0)
model.zero_grad()
if loss_str == 'l2':
loss = torch.pow(w - input_var, 2).sum()
elif loss_str == 'linf':
loss = torch.clamp((w - input_var).abs() - bound, min=0).sum()
else:
raise ValueError('Unsupported loss: %s' % loss_str)
recons_loss = loss.data.item()
w_data = w.data
if crop_frac < 1 and i % 3 == 1:
w_cropped = torch.zeros(w.size(0), w.size(1),
int(crop_frac * w.size(2)),
int(crop_frac * w.size(3)))
locs = torch.zeros(w.size(0), 4).long()
w_in = torch.zeros(w.size())
for m in range(w.size(0)):
locs[m] = torch.LongTensor(
Crop('random', crop_frac)(w_data[m]))
w_cropped = w_data[m, :, locs[m][0]:(locs[m][0] + locs[m][2]),
locs[m][1]:(locs[m][1] + locs[m][3])]
minimum = w_cropped.min()
maximum = w_cropped.max() - minimum
w_in[m] = to_tensor(
scale_up(to_pil((w_cropped - minimum) / maximum)))
w_in[m] = w_in[m] * maximum + minimum
w_in = torch.autograd.Variable(w_in, requires_grad=True)
else:
w_in = torch.autograd.Variable(w_data, requires_grad=True)
if drop_rate == 0 and i % 3 == 2:
output = model.forward(w_in)
else:
output = model.forward(torch.nn.Dropout(p=drop_rate).forward(w_in))
for j in range(output.size(0)):
loss += weight * torch.clamp(
output[j][target[j]] - output[j][argmax[j]] + kappa, min=0)
adv_loss = loss.data.item() - recons_loss
if is_gpu:
loss = loss.cuda()
loss.backward()
if crop_frac < 1 and i % 3 == 1:
grad_full = torch.zeros(w.size())
grad_cpu = w_in.grad.data
for m in range(w.size(0)):
minimum = grad_cpu[m].min()
maximum = grad_cpu[m].max() - minimum
grad_m = to_tensor(
scale_down(to_pil((grad_cpu[m] - minimum) / maximum)))
grad_m = grad_m * maximum + minimum
grad_full[m, :, locs[m][0]:(locs[m][0] + locs[m][2]),
locs[m][1]:(locs[m][1] + locs[m][3])] = grad_m
w.grad.data.add_(grad_full)
else:
w.grad.data.add_(w_in.grad.data)
w_cpu = w.data.cpu().numpy()
input_np = input.cpu().numpy()
tv_loss = 0
if tv_weight > 0:
for j in range(output.size(0)):
for k in range(3):
tv_loss += tv_weight * minimize_tv.tv(
w_cpu[j, k] - input_np[j, k], p)
grad = tv_weight * torch.from_numpy(
minimize_tv.tv_dx(w_cpu[j, k] - input_np[j, k], p))
w.grad.data[j, k].add_(grad.float())
optimizer.step()
total_loss = loss.data.cpu().item() + tv_loss
# w.data = utils.img_to_tensor(utils.transform_img(w.data), scale=False)
output_vec = w.data
preds = util.get_labels(model, output_vec)
output_vec = output_vec.view(output_vec.size(0), -1)
diff = (input_vec - output_vec).norm(2, 1).squeeze()
diff = diff.div(input_vec.norm(2, 1).squeeze())
rb = diff.mean()
sr = float(preds.ne(target).sum()) / target.size(0)
if verbose:
print(
'iteration %d: loss = %f, %s_loss = %f, '
'adv_loss = %f, tv_loss = %f' %
(i + 1, total_loss, loss_str, recons_loss, adv_loss, tv_loss))
print('robustness = %f, success rate = %f' % (rb, sr))
if total_loss < best_loss:
best_loss = total_loss
best_w = w.data.clone()
pred_xp = util.get_labels(model, best_w)
status = torch.zeros(input.size(0)).long()
status[corr] = 2 * pred[corr].ne(pred_xp[corr]).long() - 1
return (status, best_w)