surrogate_model.cuda()
surrogate_model.eval()
attacker = SwitchNeg(args.dataset, args.batch_size, args.targeted,
args.target_type, args.epsilon, args.norm, 0.0, 1.0,
args.max_queries)
for arch in archs:
if args.attack_defense:
save_result_path = args.exp_dir + "/{}_{}_result.json".format(
arch, args.defense_model)
else:
save_result_path = args.exp_dir + "/{}_result.json".format(arch)
if os.path.exists(save_result_path):
continue
log.info("Begin attack {} on {}, result will be saved to {}".format(
arch, args.dataset, save_result_path))
if args.attack_defense:
model = DefensiveModel(args.dataset,
arch,
no_grad=True,
defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda()
model.eval()
attacker.attack_all_images(args, arch, model, surrogate_model,
save_result_path)
model.cpu()
log.info("Save result of attacking {} done".format(arch))
def main():
parser = argparse.ArgumentParser(
description='Square Attack Hyperparameters.')
parser.add_argument('--norm',
type=str,
required=True,
choices=['l2', 'linf'])
parser.add_argument('--dataset', type=str, required=True)
parser.add_argument('--exp-dir',
default='logs',
type=str,
help='directory to save results and logs')
parser.add_argument(
'--gpu',
type=str,
required=True,
help='GPU number. Multiple GPUs are possible for PT models.')
parser.add_argument(
'--p',
type=float,
default=0.05,
help=
'Probability of changing a coordinate. Note: check the paper for the best values. '
'Linf standard: 0.05, L2 standard: 0.1. But robust models require higher p.'
)
parser.add_argument('--epsilon', type=float, help='Radius of the Lp ball.')
parser.add_argument('--max_queries', type=int, default=10000)
parser.add_argument(
'--json-config',
type=str,
default=
'/home1/machen/meta_perturbations_black_box_attack/configures/square_attack_conf.json',
help='a configures file to be passed in instead of arguments')
parser.add_argument('--batch_size', type=int, default=100)
parser.add_argument('--targeted', action="store_true")
parser.add_argument('--target_type',
type=str,
default='increment',
choices=['random', 'least_likely', "increment"])
parser.add_argument('--attack_defense', action="store_true")
parser.add_argument('--defense_model', type=str, default=None)
parser.add_argument('--arch',
default=None,
type=str,
help='network architecture')
parser.add_argument('--test_archs', action="store_true")
args = parser.parse_args()
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
if args.json_config:
# If a json file is given, use the JSON file as the base, and then update it with args
defaults = json.load(open(args.json_config))[args.dataset][args.norm]
arg_vars = vars(args)
arg_vars = {
k: arg_vars[k]
for k in arg_vars if arg_vars[k] is not None
}
defaults.update(arg_vars)
args = SimpleNamespace(**defaults)
if args.targeted and args.dataset == "ImageNet":
args.max_queries = 50000
args.exp_dir = os.path.join(
args.exp_dir,
get_exp_dir_name(args.dataset, args.norm, args.targeted,
args.target_type, args))
os.makedirs(args.exp_dir, exist_ok=True)
if args.test_archs:
if args.attack_defense:
log_file_path = osp.join(
args.exp_dir, 'run_defense_{}.log'.format(args.defense_model))
else:
log_file_path = osp.join(args.exp_dir, 'run.log')
elif args.arch is not None:
if args.attack_defense:
log_file_path = osp.join(
args.exp_dir,
'run_defense_{}_{}.log'.format(args.arch, args.defense_model))
else:
log_file_path = osp.join(args.exp_dir,
'run_{}.log'.format(args.arch))
set_log_file(log_file_path)
if args.test_archs:
archs = []
if args.dataset == "CIFAR-10" or args.dataset == "CIFAR-100":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/{}/checkpoint.pth.tar".format(
PY_ROOT, args.dataset, arch)
if os.path.exists(test_model_path):
archs.append(arch)
else:
log.info(test_model_path + " does not exists!")
elif args.dataset == "TinyImageNet":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{root}/train_pytorch_model/real_image_model/{dataset}@{arch}*.pth.tar".format(
root=PY_ROOT, dataset=args.dataset, arch=arch)
test_model_path = list(glob.glob(test_model_list_path))
if test_model_path and os.path.exists(test_model_path[0]):
archs.append(arch)
else:
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/checkpoints/{}*.pth".format(
PY_ROOT, args.dataset, arch)
test_model_list_path = list(glob.glob(test_model_list_path))
if len(test_model_list_path
) == 0: # this arch does not exists in args.dataset
continue
archs.append(arch)
else:
assert args.arch is not None
archs = [args.arch]
args.arch = ", ".join(archs)
log.info('Command line is: {}'.format(' '.join(sys.argv)))
log.info("Log file is written in {}".format(log_file_path))
log.info('Called with args:')
print_args(args)
attacker = MetaSimulatorSquareAttack(args.dataset,
args.batch_size,
args.targeted,
args.target_type,
args.epsilon,
args.norm,
max_queries=args.max_queries)
for arch in archs:
if args.attack_defense:
save_result_path = args.exp_dir + "/{}_{}_result.json".format(
arch, args.defense_model)
else:
save_result_path = args.exp_dir + "/{}_result.json".format(arch)
if os.path.exists(save_result_path):
continue
log.info("Begin attack {} on {}, result will be saved to {}".format(
arch, args.dataset, save_result_path))
if args.attack_defense:
model = DefensiveModel(args.dataset,
arch,
no_grad=True,
defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda()
model.eval()
attacker.attack_all_images(args, arch, model, save_result_path)
def main():
args = get_parse_args()
if args.json_config:
# If a json file is given, use the JSON file as the base, and then update it with args
defaults = json.load(open(args.json_config))[args.dataset][args.norm]
arg_vars = vars(args)
arg_vars = {k: arg_vars[k] for k in arg_vars if arg_vars[k] is not None}
defaults.update(arg_vars)
args = SimpleNamespace(**defaults)
args.exp_dir = os.path.join(args.exp_dir,
get_exp_dir_name(args.dataset, args.norm, args.targeted, args.target_type, args))
os.makedirs(args.exp_dir, exist_ok=True)
if args.test_archs:
if args.attack_defense:
log_file_path = os.path.join(args.exp_dir, 'run_defense_{}.log'.format(args.defense_model))
else:
log_file_path = os.path.join(args.exp_dir, 'run.log')
elif args.arch is not None:
if args.attack_defense:
log_file_path = os.path.join(args.exp_dir, 'run_defense_{}_{}.log'.format(args.arch, args.defense_model))
else:
log_file_path = os.path.join(args.exp_dir, 'run_{}.log'.format(args.arch))
set_log_file(log_file_path)
if args.attack_defense:
assert args.defense_model is not None
attacker = FrankWolfeBlackBoxAttack(args, args.dataset, args.targeted, args.target_type, args.epsilon,
args.norm, args.sensing, args.grad_est, args.delta, 0, 1, max_queries=args.max_queries)
torch.backends.cudnn.deterministic = True
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.test_archs:
archs = []
if args.dataset == "CIFAR-10" or args.dataset == "CIFAR-100":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/{}/checkpoint.pth.tar".format(
PY_ROOT,
args.dataset, arch)
if os.path.exists(test_model_path):
archs.append(arch)
else:
log.info(test_model_path + " does not exists!")
elif args.dataset == "TinyImageNet":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{root}/train_pytorch_model/real_image_model/{dataset}@{arch}*.pth.tar".format(
root=PY_ROOT, dataset=args.dataset, arch=arch)
test_model_path = list(glob.glob(test_model_list_path))
if test_model_path and os.path.exists(test_model_path[0]):
archs.append(arch)
else:
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/checkpoints/{}*.pth".format(
PY_ROOT,
args.dataset, arch)
test_model_list_path = list(glob.glob(test_model_list_path))
if len(test_model_list_path) == 0: # this arch does not exists in args.dataset
continue
archs.append(arch)
else:
assert args.arch is not None
archs = [args.arch]
args.arch = ", ".join(archs)
log.info('Command line is: {}'.format(' '.join(sys.argv)))
log.info("Log file is written in {}".format(log_file_path))
log.info('Called with args:')
print_args(args)
for arch in archs:
if args.attack_defense:
save_result_path = args.exp_dir + "/{}_{}_result.json".format(arch, args.defense_model)
else:
save_result_path = args.exp_dir + "/{}_result.json".format(arch)
if os.path.exists(save_result_path):
continue
log.info("Begin attack {} on {}, result will be saved to {}".format(arch, args.dataset, save_result_path))
if args.attack_defense:
model = DefensiveModel(args.dataset, arch, no_grad=True, defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda()
model.eval()
attacker.attack_all_images(args, arch, model, save_result_path)
model.cpu()
def main():
args = get_args_parse()
os.environ[
"TORCH_HOME"] = "/home1/machen/meta_perturbations_black_box_attack/train_pytorch_model/real_image_model/ImageNet-pretrained"
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
target_str = "targeted" if args.targeted else "untargeted"
json_conf = json.load(open(
args.json_config))[args.dataset][target_str][args.norm]
args = vars(args)
args.update(json_conf)
args = SimpleNamespace(**args)
if args.targeted:
if args.dataset == "ImageNet":
args.max_queries = 50000
args.exp_dir = osp.join(args.exp_dir,
get_exp_dir_name(args.dataset, args.surrogate_arch,
args.norm, args.targeted,
args.target_type,
args)) # 随机产生一个目录用于实验
os.makedirs(args.exp_dir, exist_ok=True)
if args.test_archs:
if args.attack_defense:
log_file_path = osp.join(
args.exp_dir, 'run_defense_{}.log'.format(args.defense_model))
else:
log_file_path = osp.join(args.exp_dir, 'run.log')
elif args.arch is not None:
if args.attack_defense:
log_file_path = osp.join(
args.exp_dir,
'run_defense_{}_{}.log'.format(args.arch, args.defense_model))
else:
log_file_path = osp.join(args.exp_dir,
'run_{}.log'.format(args.arch))
set_log_file(log_file_path)
archs = get_model_names(args)
args.arch = ", ".join(archs)
log.info('Command line is: {}'.format(' '.join(sys.argv)))
log.info("Log file is written in {}".format(log_file_path))
log.info('Called with args:')
print_args(args)
layer = ['fc']
extractors = []
if args.surrogate_arch == "resnet50":
resnet50 = models.resnet50(pretrained=True).eval()
resnet50_extractor = ResNetFeatureExtractor(resnet50,
layer).eval().cuda()
extractors.append(resnet50_extractor)
elif args.surrogate_arch == "resnet101":
resnet101 = models.resnet101(pretrained=True).eval()
resnet101_extractor = ResNetFeatureExtractor(resnet101,
layer).eval().cuda()
extractors.append(resnet101_extractor)
elif args.surrogate_arch == "densenet121":
densenet121 = models.densenet121(pretrained=True).eval()
densenet121_extractor = DensenetFeatureExtractor(densenet121,
layer).eval().cuda()
extractors.append(densenet121_extractor)
elif args.surrogate_arch == "densenet169":
densenet169 = models.densenet169(pretrained=True).eval()
densenet169_extractor = DensenetFeatureExtractor(densenet169,
layer).eval().cuda()
extractors.append(densenet169_extractor)
directions_generator = TentativePerturbationGenerator(extractors,
norm=args.norm,
part_size=32,
preprocess=True)
attacker = VBADAttack(args, directions_generator)
for arch in archs:
if args.attack_defense:
save_result_path = args.exp_dir + "/{}_{}_result.json".format(
arch, args.defense_model)
else:
save_result_path = args.exp_dir + "/{}_result.json".format(arch)
if os.path.exists(save_result_path):
continue
if args.attack_defense:
model = DefensiveModel(args.dataset,
arch,
no_grad=True,
defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda()
model.eval()
attacker.attack_all_images(args, arch, model, save_result_path)
model.cpu()
def attack_all_images(self, args, arch, tmp_dump_path, result_dump_path):
# subset_pos用于回调函数汇报汇总统计结果
if args.attack_defense:
model = DefensiveModel(args.dataset,
arch,
no_grad=True,
defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda()
model.eval()
# 带有缩减功能的，攻击成功的图片自动删除掉
for data_idx, data_tuple in enumerate(self.dataset_loader):
if os.path.exists(tmp_dump_path):
with open(tmp_dump_path, "r") as file_obj:
json_content = json.load(file_obj)
resume_batch_idx = int(json_content["batch_idx"]) # resume
for key in [
'query_all', 'correct_all', 'not_done_all',
'success_all', 'success_query_all'
]:
if key in json_content:
setattr(
self, key,
torch.from_numpy(np.asarray(
json_content[key])).float())
if data_idx < resume_batch_idx: # resume
continue
if args.dataset == "ImageNet":
if model.input_size[-1] >= 299:
images, true_labels = data_tuple[1], data_tuple[2]
else:
images, true_labels = data_tuple[0], data_tuple[2]
else:
images, true_labels = data_tuple[0], data_tuple[1]
if images.size(-1) != model.input_size[-1]:
images = F.interpolate(images,
size=model.input_size[-1],
mode='bilinear',
align_corners=True)
# skip_batch_index_list = np.nonzero(np.asarray(chunk_skip_indexes[data_idx]))[0].tolist()
selected = torch.arange(
data_idx * args.batch_size,
min((data_idx + 1) * args.batch_size,
self.total_images)) # 选择这个batch的所有图片的index
img_idx_to_batch_idx = ImageIdxToOrigBatchIdx(args.batch_size)
# if len(skip_batch_index_list) > 0: # FIXME 一种用bandits失败的结果就不判断的加速手段，可以删去
# for skip_index in skip_batch_index_list:
# pos = selected[skip_index]
# self.query_all[pos] = args.max_queries
# self.correct_all[pos] = 0
# self.not_done_all[pos] = 1
# self.success_all[pos] = 0 # 让其定义为分类失败
# self.success_query_all[pos] = 0
# self.not_done_loss_all[pos] = 1.0
# self.not_done_prob_all[pos] = 1.0
# images, true_labels = self.delete_tensor_by_index_list(skip_batch_index_list, images, true_labels)
# img_idx_to_batch_idx.del_by_index_list(skip_batch_index_list)
images, true_labels = images.cuda(), true_labels.cuda()
first_finetune = True
finetune_queue = FinetuneQueue(args.batch_size, args.meta_seq_len,
img_idx_to_batch_idx)
prior_size = model.input_size[
-1] if not args.tiling else args.tile_size
assert args.tiling == (args.dataset == "ImageNet")
if args.tiling:
upsampler = Upsample(size=(model.input_size[-2],
model.input_size[-1]))
else:
upsampler = lambda x: x
with torch.no_grad():
logit = model(images)
pred = logit.argmax(dim=1)
query = torch.zeros(images.size(0)).cuda()
correct = pred.eq(true_labels).float() # shape = (batch_size,)
not_done = correct.clone() # shape = (batch_size,)
if args.targeted:
if args.target_type == 'random':
target_labels = torch.randint(
low=0,
high=CLASS_NUM[args.dataset],
size=true_labels.size()).long().cuda()
invalid_target_index = target_labels.eq(true_labels)
while invalid_target_index.sum().item() > 0:
target_labels[invalid_target_index] = torch.randint(
low=0,
high=logit.shape[1],
size=target_labels[invalid_target_index].shape
).long().cuda()
invalid_target_index = target_labels.eq(true_labels)
elif args.target_type == 'least_likely':
target_labels = logit.argmin(dim=1)
elif args.target_type == "increment":
target_labels = torch.fmod(true_labels + 1,
CLASS_NUM[args.dataset])
else:
raise NotImplementedError('Unknown target_type: {}'.format(
args.target_type))
else:
target_labels = None
prior = torch.zeros(images.size(0), IN_CHANNELS[args.dataset],
prior_size, prior_size).cuda()
prior_step = self.gd_prior_step if args.norm == 'l2' else self.eg_prior_step
image_step = self.l2_image_step if args.norm == 'l2' else self.linf_step
proj_step = self.l2_proj_step if args.norm == 'l2' else self.linf_proj_step # 调用proj_maker返回的是一个函数
criterion = self.cw_loss if args.data_loss == "cw" else self.xent_loss
adv_images = images.clone()
for step_index in range(1, args.max_queries + 1):
# Create noise for exporation, estimate the gradient, and take a PGD step
dim = prior.nelement() / images.size(
0) # nelement() --> total number of elements
exp_noise = args.exploration * torch.randn_like(prior) / (
dim**0.5
) # parameterizes the exploration to be done around the prior
exp_noise = exp_noise.cuda()
q1 = upsampler(
prior + exp_noise
) # 这就是Finite Difference算法， prior相当于论文里的v，这个prior也会更新，把梯度累积上去
q2 = upsampler(
prior -
exp_noise) # prior 相当于累积的更新量，用这个更新量，再去修改image，就会变得非常准
# Loss points for finite difference estimator
q1_images = adv_images + args.fd_eta * q1 / self.norm(q1)
q2_images = adv_images + args.fd_eta * q2 / self.norm(q2)
predict_by_target_model = False
if (step_index <= args.warm_up_steps or (
step_index - args.warm_up_steps) % args.meta_predict_steps == 0) \
or (len(np.where(not_done.detach().cpu().numpy().astype(np.int32) == 1)[0]) / float(args.batch_size) <= args.notdone_threshold):
log.info("predict from target model")
predict_by_target_model = True
with torch.no_grad():
q1_logits = model(q1_images)
q2_logits = model(q2_images)
q1_logits = q1_logits / torch.norm(
q1_logits, p=2, dim=-1,
keepdim=True) # 加入normalize
q2_logits = q2_logits / torch.norm(
q2_logits, p=2, dim=-1, keepdim=True)
finetune_queue.append(q1_images.detach(),
q2_images.detach(),
q1_logits.detach(),
q2_logits.detach())
if (step_index >= args.warm_up_steps and len(
np.where(not_done.detach().cpu().numpy().astype(
np.int32) == 1)[0]) / float(args.batch_size) >
args.notdone_threshold):
q1_images_seq, q2_images_seq, q1_logits_seq, q2_logits_seq = finetune_queue.stack_history_track(
)
finetune_times = args.finetune_times if first_finetune else random.randint(
3, 5) # FIXME
self.meta_finetuner.finetune(
q1_images_seq, q2_images_seq, q1_logits_seq,
q2_logits_seq, finetune_times, first_finetune,
img_idx_to_batch_idx)
first_finetune = False
else:
with torch.no_grad():
q1_logits, q2_logits = self.meta_finetuner.predict(
q1_images, q2_images, img_idx_to_batch_idx)
q1_logits = q1_logits / torch.norm(
q1_logits, p=2, dim=-1, keepdim=True)
q2_logits = q2_logits / torch.norm(
q2_logits, p=2, dim=-1, keepdim=True)
l1 = criterion(q1_logits, true_labels, target_labels)
l2 = criterion(q2_logits, true_labels, target_labels)
# Finite differences estimate of directional derivative
est_deriv = (l1 - l2) / (args.fd_eta * args.exploration
) # 方向导数 , l1和l2是loss
# 2-query gradient estimate
est_grad = est_deriv.view(-1, 1, 1,
1) * exp_noise # B, C, H, W,
# Update the prior with the estimated gradient
prior = prior_step(
prior, est_grad,
args.online_lr) # 注意，修正的是prior,这就是bandit算法的精髓
grad = upsampler(prior) # prior相当于梯度
## Update the image:
adv_images = image_step(
adv_images,
grad * correct.view(-1, 1, 1, 1), # 注意correct也是删减过的
args.image_lr) # prior放大后相当于累积的更新量，可以用来更新
adv_images = proj_step(images, args.epsilon, adv_images)
adv_images = torch.clamp(adv_images, 0, 1)
with torch.no_grad():
adv_logit = model(adv_images) #
adv_pred = adv_logit.argmax(dim=1)
adv_prob = F.softmax(adv_logit, dim=1)
adv_loss = criterion(adv_logit, true_labels, target_labels)
## Continue query count
if predict_by_target_model:
query = query + 2 * not_done
if args.targeted:
not_done = not_done * (
1 - adv_pred.eq(target_labels).float()
).float() # not_done初始化为 correct, shape = (batch_size,)
else:
not_done = not_done * adv_pred.eq(
true_labels).float() # 只要是跟原始label相等的，就还需要query，还没有成功
success = (1 - not_done) * correct
success_query = success * query
not_done_loss = adv_loss * not_done
not_done_prob = adv_prob[torch.arange(adv_images.size(0)),
true_labels] * not_done
log.info('Attacking image {} - {} / {}, step {}'.format(
data_idx * args.batch_size,
(data_idx + 1) * args.batch_size, self.total_images,
step_index))
log.info(' not_done: {:.4f}'.format(
len(
np.where(not_done.detach().cpu().numpy().astype(
np.int32) == 1)[0]) / float(args.batch_size)))
log.info(' fd_scalar: {:.9f}'.format(
(l1 - l2).mean().item()))
if success.sum().item() > 0:
log.info(' mean_query: {:.4f}'.format(
success_query[success.byte()].mean().item()))
log.info(' median_query: {:.4f}'.format(
success_query[success.byte()].median().item()))
if not_done.sum().item() > 0:
log.info(' not_done_loss: {:.4f}'.format(
not_done_loss[not_done.byte()].mean().item()))
log.info(' not_done_prob: {:.4f}'.format(
not_done_prob[not_done.byte()].mean().item()))
not_done_np = not_done.detach().cpu().numpy().astype(np.int32)
done_img_idx_list = np.where(not_done_np == 0)[0].tolist()
delete_all = False
if done_img_idx_list:
for skip_index in done_img_idx_list: # 两次循环，第一次循环先汇报出去，第二次循环删除
batch_idx = img_idx_to_batch_idx[skip_index]
pos = selected[batch_idx].item()
# 先汇报被删减的值self.query_all
for key in [
'query', 'correct', 'not_done', 'success',
'success_query', 'not_done_loss',
'not_done_prob'
]:
value_all = getattr(self, key + "_all")
value = eval(key)[skip_index].item()
value_all[pos] = value
images, adv_images, prior, query, true_labels, target_labels, correct, not_done =\
self.delete_tensor_by_index_list(done_img_idx_list, images, adv_images, prior, query, true_labels, target_labels, correct, not_done)
img_idx_to_batch_idx.del_by_index_list(done_img_idx_list)
delete_all = images is None
if delete_all:
break
# report to all stats the rest unsuccess
for key in [
'query', 'correct', 'not_done', 'success', 'success_query',
'not_done_loss', 'not_done_prob'
]:
for img_idx, batch_idx in img_idx_to_batch_idx.proj_dict.items(
):
pos = selected[batch_idx].item()
value_all = getattr(self, key + "_all")
value = eval(key)[img_idx].item()
value_all[
pos] = value # 由于value_all是全部图片都放在一个数组里，当前batch选择出来
img_idx_to_batch_idx.proj_dict.clear()
tmp_info_dict = {
"batch_idx": data_idx + 1,
"batch_size": args.batch_size
}
for key in [
'query_all', 'correct_all', 'not_done_all', 'success_all',
'success_query_all'
]:
value_all = getattr(self, key).detach().cpu().numpy().tolist()
tmp_info_dict[key] = value_all
with open(tmp_dump_path, "w") as result_file_obj:
json.dump(tmp_info_dict, result_file_obj, sort_keys=True)
query_all_ = self.query_all.detach().cpu().numpy().astype(np.int32)
not_done_all_ = self.not_done_all.detach().cpu().numpy().astype(
np.int32)
# log.info('{} is attacked finished ({} images)'.format(arch, self.total_images))
# log.info(' avg correct: {:.4f}'.format(self.correct_all.mean().item()))
# log.info(' avg not_done: {:.4f}'.format(self.not_done_all.mean().item())) # 有多少图没做完
# if self.success_all.sum().item() > 0:
# log.info(
# ' avg mean_query: {:.4f}'.format(self.success_query_all[self.success_all.byte()].mean().item()))
# log.info(
# ' avg median_query: {:.4f}'.format(self.success_query_all[self.success_all.byte()].median().item()))
# log.info(' max query: {}'.format(self.success_query_all[self.success_all.byte()].max().item()))
# if self.not_done_all.sum().item() > 0:
# log.info(
# ' avg not_done_loss: {:.4f}'.format(self.not_done_loss_all[self.not_done_all.byte()].mean().item()))
# log.info(
# ' avg not_done_prob: {:.4f}'.format(self.not_done_prob_all[self.not_done_all.byte()].mean().item()))
log.info('Saving results to {}'.format(result_dump_path))
meta_info_dict = {
"avg_correct":
self.correct_all.mean().item(),
"avg_not_done":
self.not_done_all[self.correct_all.byte()].mean().item(),
"mean_query":
self.success_query_all[self.success_all.byte()].mean().item(),
"median_query":
self.success_query_all[self.success_all.byte()].median().item(),
"max_query":
self.success_query_all[self.success_all.byte()].max().item(),
"correct_all":
self.correct_all.detach().cpu().numpy().astype(np.int32).tolist(),
"not_done_all":
self.not_done_all.detach().cpu().numpy().astype(np.int32).tolist(),
"query_all":
self.query_all.detach().cpu().numpy().astype(np.int32).tolist(),
"not_done_loss":
self.not_done_loss_all[self.not_done_all.byte()].mean().item(),
"not_done_prob":
self.not_done_prob_all[self.not_done_all.byte()].mean().item(),
"args":
vars(args)
}
with open(result_dump_path, "w") as result_file_obj:
json.dump(meta_info_dict, result_file_obj, sort_keys=True)
log.info("done, write stats info to {}".format(result_dump_path))
self.query_all.fill_(0)
self.correct_all.fill_(0)
self.not_done_all.fill_(0)
self.success_all.fill_(0)
self.success_query_all.fill_(0)
self.not_done_loss_all.fill_(0)
self.not_done_prob_all.fill_(0)
model.cpu()
def main(args, result_dir_path):
log.info('Loading %s model and test data' % args.dataset)
dataset_loader = DataLoaderMaker.get_test_attacked_data(args.dataset, 1)
if args.test_archs:
archs = []
if args.dataset == "CIFAR-10" or args.dataset == "CIFAR-100":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/{}/checkpoint.pth.tar".format(
PY_ROOT, args.dataset, arch)
if os.path.exists(test_model_path):
archs.append(arch)
else:
log.info(test_model_path + " does not exists!")
elif args.dataset == "TinyImageNet":
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{root}/train_pytorch_model/real_image_model/{dataset}@{arch}*.pth.tar".format(
root=PY_ROOT, dataset=args.dataset, arch=arch)
test_model_path = list(glob.glob(test_model_list_path))
if test_model_path and os.path.exists(test_model_path[0]):
archs.append(arch)
else:
for arch in MODELS_TEST_STANDARD[args.dataset]:
test_model_list_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/checkpoints/{}*.pth".format(
PY_ROOT, args.dataset, arch)
test_model_list_path = list(glob.glob(test_model_list_path))
if len(test_model_list_path
) == 0: # this arch does not exists in args.dataset
continue
archs.append(arch)
else:
assert args.arch is not None
archs = [args.arch]
if args.attack_defense:
meta_model_path = '{}/train_pytorch_model/meta_grad_regression/{}_without_resnet.pth.tar'.format(
PY_ROOT, args.dataset)
else:
meta_model_path = '{}/train_pytorch_model/meta_grad_regression/{}.pth.tar'.format(
PY_ROOT, args.dataset)
assert os.path.exists(meta_model_path), "{} does not exist!".format(
meta_model_path)
meta_model = load_meta_model(meta_model_path)
log.info("Load meta model from {}".format(meta_model_path))
attack = MetaGradAttack(args,
norm=args.norm,
epsilon=args.epsilon,
targeted=args.targeted,
search_steps=args.binary_steps,
max_steps=args.maxiter,
use_log=not args.use_zvalue,
cuda=not args.no_cuda)
for arch in archs:
if args.attack_defense:
model = DefensiveModel(args.dataset,
arch,
no_grad=True,
defense_model=args.defense_model)
else:
model = StandardModel(args.dataset, arch, no_grad=True)
model.cuda().eval()
query_all = []
not_done_all = []
correct_all = []
img_no = 0
total_success = 0
l2_total = 0.0
avg_step = 0
avg_time = 0
avg_qry = 0
if args.attack_defense:
result_dump_path = result_dir_path + "/{}_{}_result.json".format(
arch, args.defense_model)
else:
result_dump_path = result_dir_path + "/{}_result.json".format(arch)
# if os.path.exists(result_dump_path):
# continue
log.info("Begin attack {} on {}".format(arch, args.dataset))
for i, data_tuple in enumerate(dataset_loader):
if args.dataset == "ImageNet":
if model.input_size[-1] >= 299:
img, true_labels = data_tuple[1], data_tuple[2]
else:
img, true_labels = data_tuple[0], data_tuple[2]
else:
img, true_labels = data_tuple[0], data_tuple[1]
args.init_size = model.input_size[-1]
if img.size(-1) != model.input_size[-1]:
img = F.interpolate(img,
size=model.input_size[-1],
mode='bilinear',
align_corners=True)
img, true_labels = img.to(0), true_labels.to(0)
with torch.no_grad():
pred_logit = model(img)
pred_label = pred_logit.argmax(dim=1)
correct = pred_label.eq(true_labels).detach().cpu().numpy().astype(
np.int32)
correct_all.append(correct)
if pred_label[0].item() != true_labels[0].item():
log.info(
"Skip wrongly classified image no. %d, original class %d, classified as %d"
% (i, pred_label.item(), true_labels.item()))
query_all.append(0)
not_done_all.append(
1) # 原本就分类错误，not_done给1，假如99%都原本分类错误的话, avg_not_done = 99%
continue
img_no += 1
timestart = time.time()
meta_model_copy = copy.deepcopy(meta_model)
if args.targeted:
if args.target_type == 'random':
target_labels = torch.randint(
low=0,
high=CLASS_NUM[args.dataset],
size=true_labels.size()).long().cuda()
invalid_target_index = target_labels.eq(true_labels)
while invalid_target_index.sum().item() > 0:
target_labels[invalid_target_index] = torch.randint(
low=0,
high=pred_logit.shape[1],
size=target_labels[invalid_target_index].shape
).long().cuda()
invalid_target_index = target_labels.eq(true_labels)
elif args.target_type == 'least_likely':
target_labels = pred_logit.argmin(dim=1)
elif args.target_type == "increment":
target_labels = torch.fmod(true_labels + 1,
CLASS_NUM[args.dataset])
else:
raise NotImplementedError('Unknown target_type: {}'.format(
args.target_type))
else:
target_labels = None
target = true_labels if not args.targeted else target_labels
adv, const, first_step, success_queries = attack.run(
model, meta_model_copy, img, target)
timeend = time.time()
if len(adv.shape) == 3:
adv = adv.reshape((1, ) + adv.shape)
adv = torch.from_numpy(adv).permute(0, 3, 1,
2).cuda() # BHWC -> BCHW
diff = (adv - img).detach().cpu().numpy()
l2_distortion = np.sqrt(np.sum(np.square(diff))).item()
with torch.no_grad():
adv_pred_logit = model(adv)
adv_pred_label = adv_pred_logit.argmax(dim=1)
success = False
if not args.targeted: # target is true label
if adv_pred_label[0].item() != target[0].item():
success = True
else:
if adv_pred_label[0].item() == target[0].item():
success = True
if success_queries > args.max_queries:
success = False
if success:
# (first_step-1)//args.finetune_intervalargs.update_pixels2+first_step
# The first step is the iterations used that find the adversarial examples;
# args.finetune_interval is the finetuning per iterations;
# args.update_pixels is the queried pixels each iteration.
# Currently, we find only f(x+h)-f(x) could estimate the gradient well, so args.update_pixels*1 in my updated codes.
not_done_all.append(0)
# only 1 query for i pixle, because the estimated function is f(x+h)-f(x)/h
query_all.append(success_queries)
total_success += 1
l2_total += l2_distortion
avg_step += first_step
avg_time += timeend - timestart
# avg_qry += (first_step-1)//args.finetune_interval*args.update_pixels*1+first_step
avg_qry += success_queries
log.info("Attack {}-th image: {}, query:{}".format(
i, "success", success_queries))
else:
not_done_all.append(1)
query_all.append(args.max_queries)
log.info("Attack {}-th image: {}, query:{}".format(
i, "fail", success_queries))
model.cpu()
if total_success != 0:
log.info(
"[STATS][L1] total = {}, time = {:.3f}, distortion = {:.5f}, avg_step = {:.5f},avg_query = {:.5f}, success_rate = {:.3f}"
.format(img_no, avg_time / total_success,
l2_total / total_success, avg_step / total_success,
avg_qry / total_success,
total_success / float(img_no)))
correct_all = np.concatenate(correct_all, axis=0).astype(np.int32)
query_all = np.array(query_all).astype(np.int32)
not_done_all = np.array(not_done_all).astype(np.int32)
success = (1 - not_done_all) * correct_all
success_query = success * query_all
# query_all_bounded = query_all.copy()
# not_done_all_bounded = not_done_all.copy()
# out_of_bound_indexes = np.where(query_all_bounded > args.max_queries)[0]
# if len(out_of_bound_indexes) > 0:
# not_done_all_bounded[out_of_bound_indexes] = 1
# success_bounded = (1-not_done_all_bounded) * correct_all
# success_query_bounded = success_bounded * query_all_bounded
#
# query_threshold_success_rate_bounded, query_success_rate_bounded = success_rate_and_query_coorelation(query_all_bounded, not_done_all_bounded)
# success_rate_to_avg_query_bounded = success_rate_avg_query(query_all_bounded, not_done_all_bounded)
meta_info_dict = {
"query_all":
query_all.tolist(),
"not_done_all":
not_done_all.tolist(),
"correct_all":
correct_all.tolist(),
"mean_query":
np.mean(success_query[np.nonzero(success)[0]]).item(),
"max_query":
np.max(success_query[np.nonzero(success)[0]]).item(),
"median_query":
np.median(success_query[np.nonzero(success)[0]]).item(),
"avg_not_done":
np.mean(
not_done_all.astype(
np.float32)[np.nonzero(correct_all)[0]]).item(),
# "mean_query_bounded_max_queries": np.mean(success_query_bounded[np.nonzero(success_bounded)[0]]).item(),
# "max_query_bounded_max_queries": np.max(success_query_bounded[np.nonzero(success_bounded)[0]]).item(),
# "median_query_bounded_max_queries": np.median(success_query_bounded[np.nonzero(success_bounded)[0]]).item(),
# "avg_not_done_bounded_max_queries": np.mean(not_done_all_bounded.astype(np.float32)).item(),
# "query_threshold_success_rate_dict_bounded": query_threshold_success_rate_bounded,
# "query_success_rate_dict_bounded": query_success_rate_bounded,
# "success_rate_to_avg_query_bounded": success_rate_to_avg_query_bounded,
"args":
vars(args)
}
with open(result_dump_path, "w") as result_file_obj:
json.dump(meta_info_dict, result_file_obj, sort_keys=True)
log.info("done, write stats info to {}".format(result_dump_path))
