def calc_cosine_sim(task, pgen_type, args, gt_dy, n_epoch=10):
B = gt_dy.shape[0]
gt_dy = torch.FloatTensor(gt_dy)
p_gen = attack_setting.load_pgen(task=task, pgen_type=pgen_type, args=args)
input_size = attack_setting.pgen_input_size(task=task,
pgen_type=pgen_type,
args=args)
latent_data = torch.randn((B, *input_size))
if args.use_gpu:
latent_data = latent_data.cuda()
gt_dy = gt_dy.cuda()
# Puts the tensor on cuda before setting requires_grad to True
# Otherwise it would generate "ValueError: can't optimize a non-leaf Tensor"
latent_data.requires_grad = True
cos_criterion = torch.nn.CosineSimilarity()
optimizer = torch.optim.Adam([latent_data], lr=1e-3)
old_loss = -999
smallest_loss = 999
with tqdm(range(n_epoch)) as pbar:
for i in pbar:
optimizer.zero_grad()
projected_data = p_gen.project(latent_data)
proj_loss = torch.mean(
-cos_criterion(projected_data.view(B, -1), gt_dy.view(B, -1)))
proj_loss.backward()
old_loss = proj_loss.detach().cpu().numpy()
if old_loss < smallest_loss:
smallest_loss = old_loss
optimizer.step()
pbar.set_description("Epoch %d, cosine similarity %f" %
(i, -old_loss))
return -smallest_loss
#rvs = p_gen.generate_ps(src_image, 10, level=3)
#t2 = time.time()
#print (rvs.shape)
#grad_gt = fmodel.gradient_one(src_image, label=src_label)
#print (grad_gt.shape)
#print (p_gen.calc_rho(grad_gt, src_image))
#t3 = time.time()
#print (t2-t1,t3-t2)
#assert 0
#for PGEN in ['PCA9408basis', 'DCT9408', 'naive', 'resize9408']:
#for PGEN in ['DCT9408', 'naive', 'resize9408']:
for PGEN in [
'naive',
]:
p_gen = load_pgen(TASK, PGEN, args)
if TASK == 'cifar':
if PGEN == 'naive':
ITER = 150
maxN = 30
initN = 30
elif PGEN.startswith('DCT') or PGEN.startswith('resize'):
ITER = 150
maxN = 30
initN = 30
elif PGEN.startswith('PCA'):
ITER = 150
maxN = 30
initN = 30
else:
raise NotImplementedError()
def calc_internal_dim(task, pgen_type, args):
B = 5000
BATCH_SIZE = 32
p_gen = attack_setting.load_pgen(task=task, pgen_type=pgen_type, args=args)
input_size = attack_setting.pgen_input_size(task=task,
pgen_type=pgen_type,
args=args)
latent_data = torch.randn((B, *input_size))
if p_gen is not None:
if args.use_gpu:
latent_data = latent_data.cuda()
# projected_data = p_gen.project(latent_data)
projected_np = None
for _i in range(int(B / BATCH_SIZE) + 1):
_data = latent_data[_i * BATCH_SIZE:(_i + 1) * BATCH_SIZE]
_B = _data.shape[0]
if _B < 1:
break
_projected = p_gen.project(_data)
_np = _projected.detach().cpu().numpy().reshape(_B, -1)
if projected_np is None:
projected_np = _np
else:
projected_np = np.concatenate((projected_np, _np), axis=0)
else:
projected_np = latent_data.numpy().reshape(B, -1)
model_file_name = model_settings.get_model_file_name(TASK=task, args=args)
if args.do_svd:
print('Doing svd...')
u, s, v = np.linalg.svd(projected_np, full_matrices=False)
np.save(
'BAPP_result/%s_%s_internal_dim_u.npy' %
(model_file_name, pgen_type), u)
np.save(
'BAPP_result/%s_%s_internal_dim_s.npy' %
(model_file_name, pgen_type), s)
np.save(
'BAPP_result/%s_%s_internal_dim_v.npy' %
(model_file_name, pgen_type), v)
else:
u = np.load('BAPP_result/%s_%s_internal_dim_u.npy' %
(model_file_name, pgen_type))
s = np.load('BAPP_result/%s_%s_internal_dim_s.npy' %
(model_file_name, pgen_type))
v = np.load('BAPP_result/%s_%s_internal_dim_v.npy' %
(model_file_name, pgen_type))
projected_np = u.dot(np.diag(s)).dot(v)
cos_sims = []
s_keep = np.zeros(s.shape)
with tqdm(range(s.shape[0])) as pbar:
for i in pbar:
s_keep[i] = s[i]
if i % inter_gap == 0:
slice = u.dot(np.diag(s_keep)).dot(v)
cos_sim = np.mean(
utils.calc_cos_sim(x1=projected_np, x2=slice, dim=1))
cos_sims.append(cos_sim)
pbar.set_description('Keep dim %d, cosine similarity %f' %
(i, cos_sim))
np.save(
'BAPP_result/%s_%s_internal_dim.npy' %
(model_file_name, pgen_type), cos_sims)
if cos_sim > 0.9999:
break
print("%s, %s, Cosine similarity for internal dim done" %
(model_file_name, pgen_type))
print(cos_sims)
def calc_exp_cos(task, pgen_type, args):
p_gen = attack_setting.load_pgen(task=task, pgen_type=pgen_type, args=args)
input_size = attack_setting.pgen_input_size(task=task,
pgen_type=pgen_type,
args=args)
REF = 'res18'
if task == 'mnist':
img_size = args.mnist_img_size
elif task == 'cifar10':
img_size = args.cifar10_img_size
else:
img_size = 224
if args.mounted:
data_dir = 'ANONYMIZED_DIRECTORY/%s_%s' % (task, REF)
if task == 'cifar10' and args.cifar10_img_size != 32:
img_size = args.cifar10_img_size
data_dir = 'ANONYMIZED_DIRECTORY/%s_%d_%s' % (task, img_size, REF)
if task == 'mnist' and args.mnist_img_size != 28:
img_size = args.mnist_img_size
data_dir = 'ANONYMIZED_DIRECTORY/%s_%d_%s' % (task, img_size, REF)
else:
data_dir = 'ANONYMIZED_DIRECTORY/%s_%s' % (task, REF)
if task == 'cifar10' and args.cifar10_img_size != 32:
img_size = args.cifar10_img_size
data_dir = 'ANONYMIZED_DIRECTORY/%s_%d_%s' % (task, img_size, REF)
if task == 'mnist' and args.mnist_img_size != 28:
img_size = args.mnist_img_size
data_dir = 'ANONYMIZED_DIRECTORY/%s_%d_%s' % (task, img_size, REF)
data_path = data_dir + '/test_batch_%d.npy'
cos_sims = []
if task == 'celeba':
N_test = 624
if REF == 'rnd':
N_test = 625
elif task == 'imagenet':
N_test = 625
if REF == 'rnd':
N_test = 625
elif task == 'celebaid':
N_test = 3
if REF == 'rnd':
N_test = 5 * 3
elif task == 'cifar10':
N_test = 313
if REF == 'rnd':
N_test = 5 * 313
elif args.TASK == 'mnist':
N_test = 313
if REF == 'rnd':
N_test = 5 * 313
with tqdm(range(min(10000, N_test))) as pbar:
for idx in pbar:
X = np.load(data_path % (idx))
B = X.shape[0]
latent_data = torch.randn((B, *input_size))
if p_gen is not None:
if args.use_gpu:
latent_data = latent_data.cuda()
projected_data = p_gen.project(latent_data)
projected_np = projected_data.detach().cpu().numpy().reshape(
B, -1)
else:
projected_np = latent_data.numpy().reshape(B, -1)
# if np.sum((X ** 2).sum(1)) < 1e-8 or np.sum((projected_np ** 2).sum(1)) < 1e-8:
# continue
cos_sim = np.mean(
np.abs(calc_cos_sim(x1=X, x2=projected_np, dim=1)))
cos_sims.append(cos_sim)
pbar.set_description('Data # %d, cosine similarity %f' %
(idx, cos_sim))
print("%s, %s, Expected cos similarity" % (task, pgen_type))
print(cos_sims)
np.save('BAPP_result/%s_%d_%s_exp_cos.npy' % (task, img_size, pgen_type),
cos_sims)
return np.mean(cos_sims)
simi_threshold=args.threshold)
src_image = load_img(src_img_path)
tgt_image = load_img(tgt_img_path)
src_label = np.argmax(fmodel.forward_one(src_image))
tgt_label = np.argmax(fmodel.forward_one(tgt_image))
if src_label == tgt_label:
print("src tgt same prediction", src_id, tgt_id)
continue
print(
"======== # %d, src id %d, tgt id %d, src label %d, tgt label %d ========"
% (i, src_id, tgt_id, src_label, tgt_label))
mask = None
p_gen = attack_setting.load_pgen(task=TASK,
pgen_type=args.pgen,
args=args)
attack = foolbox.attacks.BAPP_custom(fmodel,
criterion=TargetClass(src_label))
adv = attack(tgt_image,
tgt_label,
starting_point=src_image,
iterations=20,
batch_size=9999,
stepsize_search='geometric_progression',
verbose=True,
unpack=False,
max_num_evals=100,
initial_num_evals=100,
internal_dtype=np.float32,