def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_index
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# dataset：将数据集名称转化为大写形式MNIST/CIFAR10
dataset = args.dataset.upper()
assert dataset == 'MNIST' or dataset == 'CIFAR10'
# 获取MNIST/CIFAR10的模型训练参数，模型，训练集，验证集和测试集
if dataset == 'MNIST':
training_parameters = MNIST_Training_Parameters
model_framework = MNISTConvNet().to(device)
batch_size = training_parameters['batch_size']
train_loader, valid_loader = get_mnist_train_validate_loader(
dir_name='../data/MNIST/',
batch_size=batch_size,
valid_size=0.1,
shuffle=True)
test_loader = get_mnist_test_loader(dir_name='../data/MNIST/',
batch_size=batch_size)
else:
training_parameters = CIFAR10_Training_Parameters
model_framework = resnet20_cifar().to(device)
batch_size = training_parameters['batch_size']
train_loader, valid_loader = get_cifar10_train_validate_loader(
dir_name='../data/CIFAR10/',
batch_size=batch_size,
valid_size=0.1,
shuffle=True)
test_loader = get_cifar10_test_loader(dir_name='../data/CIFAR10/',
batch_size=batch_size)
# defense_name：防御名称为EAT
defense_name = 'NEAT'
# eat_params：防御参数
neat_params = {'eps': args.eps, 'alpha': args.alpha}
# 将各种参数传入到NEAT防御中
neat = NEATDefense(model=model_framework,
defense_name=defense_name,
dataset=dataset,
training_parameters=training_parameters,
device=device,
**neat_params)
# 如果为True则需要重新训练预定义模型
if args.train_externals:
print('\nStart to train the external models ......\n')
neat.train_external_model_group(train_loader=train_loader,
validation_loader=valid_loader)
# 加载预训练模型
pre_train_models = neat.load_external_model_group(
model_dir='../DefenseEnhancedModels/NEAT/', test_loader=test_loader)
# 进行集成对抗训练，将最佳的模型参数保存到DefenseEnhancedModels/NEAT/MNIST_NEAT_enhanced.pt中或者CIFAR10
neat.defense(pre_trained_models=pre_train_models,
train_loader=train_loader,
validation_loader=valid_loader)
def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_index
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Set the random seed manually for reproducibility.
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# 下载CIFAR10训练集，划分为训练集和测试集（然后进行分组），保存到CIFAR10文件夹下面
train_loader, valid_loader = get_cifar10_train_validate_loader(
dir_name='../data/CIFAR10/',
batch_size=CIFAR10_Training_Parameters['batch_size'],
valid_size=0.1,
shuffle=True)
# 下载CIFAR10测试集（然后进行分组），保存到CIFAR10文件夹下面
test_loader = get_cifar10_test_loader(
dir_name='../data/CIFAR10/',
batch_size=CIFAR10_Training_Parameters['batch_size'])
# 设置模型
# **************引入的模型名称**************
resnet_model = ResNet18().to(device)
# 设置优化器
optimizer = optim.Adam(resnet_model.parameters(),
lr=CIFAR10_Training_Parameters['lr'])
# 训练
# 最好的验证集精度
best_val_acc = None
# 训练模型参数保存路径：/CIFAR10/model/CIFAR10_raw.pt
# **************不同模型需要修改名称**************
model_saver = './CIFAR10/model/ResNet18_' + 'raw' + '.pt'
# 进行epoch次循环训练
for epoch in range(CIFAR10_Training_Parameters['num_epochs']):
# 一次epoch训练
train_one_epoch(model=resnet_model,
train_loader=train_loader,
optimizer=optimizer,
epoch=epoch,
device=device)
# 验证集的精度
val_acc = validation_evaluation(model=resnet_model,
validation_loader=valid_loader,
device=device)
# 学习率调整
adjust_learning_rate(optimizer=optimizer, epoch=epoch)
# 每一次epoch后验证集的精度大于最好的精度时（移除模型保存路径），或者best_val_acc为None时，更新最佳精度，然后将模型参数重新写入保存路径中
if not best_val_acc or round(val_acc, 4) >= round(best_val_acc, 4):
if best_val_acc is not None:
os.remove(model_saver)
best_val_acc = val_acc
resnet_model.save(name=model_saver)
# 否则提示精度未发生提高
else:
print(
'Train Epoch{:>3}: validation dataset accuracy did not improve from {:.4f}\n'
.format(epoch, best_val_acc))
# 测试
# 复制resnet_model
final_model = copy.deepcopy(resnet_model)
# 加载final_model
final_model.load(path=model_saver, device=device)
# 计算模型在测试集上面的精度并输出
accuracy = testing_evaluation(model=final_model,
test_loader=test_loader,
device=device)
# 打印模型在测试集上的精度
print(
'Finally, the ACCURACY of saved model [{}] on testing dataset is {:.2f}%\n'
.format(final_model.model_name, accuracy * 100.0))
def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_index
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# dataset：数据集名称MNIST或CIFAR10
dataset = args.dataset.upper()
assert dataset == 'MNIST' or dataset == 'CIFAR10'
# 加载模型和测试集
raw_model_location = '{}{}/model/{}_raw.pt'.format('../data/', dataset,
dataset)
if dataset == 'MNIST':
raw_model = MNIST_CNN().to(device)
raw_model.load(path=raw_model_location, device=device)
test_loader = get_mnist_test_loader(dir_name='../data/MNIST/',
batch_size=30)
else:
raw_model = ResNet18().to(device)
raw_model.load(path=raw_model_location, device=device)
test_loader = get_cifar10_test_loader(dir_name='../data/CIFAR10/',
batch_size=25)
raw_model.eval()
# 原始模型对测试集进行预测
predicted_raw, true_label = prediction(model=raw_model,
test_loader=test_loader,
device=device)
# 需要再训练的防御
re_train_defenses = {'NAT', 'RLT', 'RLT1', 'RLT2', 'RLT3', 'EAT', 'UAPAT'}
# 其他防御
other_defenses = {'NRC'}
# defense_name：防御名称
defense_name = args.defense.upper().strip()
# 如果是再训练模型防御
if defense_name in re_train_defenses:
print(
'\nthe ##{}## defense is a kind of complete defenses that retrain the model'
.format(defense_name))
# 加载防御模型
# defended_model_location：防御模型位置DefenseEnhancedModels/defense_name/CIFAR10_defense_name_enhanced.pt或者MNIST
defended_model_location = '{}/{}/{}_{}_enhanced.pt'.format(
'../DefenseEnhancedModels', defense_name, dataset, defense_name)
defended_model = MNIST_CNN().to(
device) if dataset == 'MNIST' else ResNet18().to(device)
defended_model.load(path=defended_model_location, device=device)
defended_model.eval()
# 利用防御模型进行标签预测
predicted_defended, _ = prediction(model=defended_model,
test_loader=test_loader,
device=device)
# 计算防御指标
raw_acc, def_acc, cav, crr, csr = defense_utility_measure(
predicted_defended, predicted_raw, true_label)
else:
if defense_name == 'NRC':
print(
'\n##{}## defense is a kind of region-based classification defenses ... '
.format(defense_name))
from Defenses.DefenseMethods.NRC import NRCDefense
num_points = 1000
radius = args.radius
mean = args.mean
std = args.std
nrc = NRCDefense(model=raw_model,
defense_name='NRC',
dataset=dataset,
device=device,
num_points=num_points)
predicted_defended = []
with torch.no_grad():
for index, (images, labels) in enumerate(test_loader):
nrc_labels = nrc.region_based_classification(
samples=images, radius=radius, mean=mean, std=std)
predicted_defended.extend(nrc_labels)
predicted_defended = np.array(predicted_defended)
correct_prediction_def = np.equal(predicted_defended, true_label)
def_acc = np.mean(correct_prediction_def.astype(float))
correct_prediction_raw = np.equal(np.argmax(predicted_raw, axis=1),
true_label)
raw_acc = np.mean(correct_prediction_raw.astype(float))
# Classification Accuracy Variance(CAV)
cav = def_acc - raw_acc
# Find the index of correct predicted examples by defence-enhanced model and raw model
idx_def = np.squeeze(np.argwhere(correct_prediction_def == True))
idx_raw = np.squeeze(np.argwhere(correct_prediction_raw == True))
idx = np.intersect1d(idx_def, idx_raw, assume_unique=True)
crr = (len(idx_def) - len(idx)) / len(predicted_raw)
csr = (len(idx_raw) - len(idx)) / len(predicted_raw)
else:
raise ValueError('{} is not supported!!!'.format(defense_name))
# 输出防御指标的值
print("****************************")
print(
"The utility evaluation results of the {} defense for {} Dataset are as follow:"
.format(defense_name, dataset))
print('Acc of Raw Model:\t\t{:.2f}%'.format(raw_acc * 100))
print('Acc of {}-enhanced Model:\t{:.2f}%'.format(defense_name,
def_acc * 100))
print('CAV: {:.2f}%'.format(cav * 100))
print('CRR: {:.2f}%'.format(crr * 100))
print('CSR: {:.2f}%'.format(csr * 100))
print("****************************")
def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_index
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Set the random seed manually for reproducibility.
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# dataset：数据集名称MNIST或CIFAR10
# num：选取的干净样本的数量
# dataset_location：数据集所在位置data/CIFAR10/或者MNIST
# raw_model_location：模型所在位置data/CIFAR10/model/CIFAR10_raw.pt或者MNIST
dataset = args.dataset.upper()
num = args.number
# *****************数据集存放的位置*****************
dataset_location = '../data/{}/'.format(dataset)
raw_model_location = '../data/{}/model/{}_raw.pt'.format(dataset, dataset)
print(
"\nStarting to select {} {} Candidates Example, which are correctly classified by the Raw Model from {}\n"
.format(num, dataset, raw_model_location))
# 加载模型，获取测试集
# raw_model：模型
# test_loader：测试集
# load the raw model and testing dataset
assert args.dataset == 'MNIST' or args.dataset == 'CIFAR10'
if dataset == 'MNIST':
raw_model = MNIST_CNN().to(device)
raw_model.load(path=raw_model_location, device=device)
test_loader = get_mnist_test_loader(dir_name=dataset_location,
batch_size=1,
shuffle=False)
else:
raw_model = ResNet18().to(device)
raw_model.load(path=raw_model_location, device=device)
test_loader = get_cifar10_test_loader(dir_name=dataset_location,
batch_size=1,
shuffle=False)
# 获取分类正确的测试集
# successful：测试集经过模型，保留被正确预测的图像和标签以及它们对应softmax最小输出的标签
successful = []
raw_model.eval()
with torch.no_grad():
for image, label in test_loader:
image = image.to(device)
label = label.to(device)
output = raw_model(image)
_, predicted = torch.max(output.data, 1)
if predicted == label:
_, least_likely_class = torch.min(output.data, 1)
successful.append([image, label, least_likely_class])
print(len(successful))
# 随机选取num个正确分类的图像
candidates = random.sample(successful, num)
candidate_images = []
candidate_labels = []
candidates_llc = []
candidate_targets = []
for index in range(len(candidates)):
# 将选择的图片，标签和最不可能的标签分开
image = candidates[index][0].cpu().numpy()
image = np.squeeze(image, axis=0)
candidate_images.append(image)
label = candidates[index][1].cpu().numpy()[0]
llc = candidates[index][2].cpu().numpy()[0]
# 生成0~9的10个标签，去除真实标签，随机选择一个标签
classes = [i for i in range(10)]
classes.remove(label)
target = random.sample(classes, 1)[0]
# 将随机目标标签，最不可能分类目标标签和真实标签转化为one-hot标签保存
one_hot_label = [0 for i in range(10)]
one_hot_label[label] = 1
one_hot_llc = [0 for i in range(10)]
one_hot_llc[llc] = 1
one_hot_target = [0 for i in range(10)]
one_hot_target[target] = 1
candidate_labels.append(one_hot_label)
candidates_llc.append(one_hot_llc)
candidate_targets.append(one_hot_target)
# 图像
candidate_images = np.array(candidate_images)
# 图像对应真实one-hot标签
candidate_labels = np.array(candidate_labels)
# 图像对应最不可能分类的one-hot标签
candidates_llc = np.array(candidates_llc)
# 图像对应非真实标签的随机one-hot标签
candidate_targets = np.array(candidate_targets)
# 打开CIFAR10/或MNIST/文件夹
if dataset not in os.listdir('./'):
os.mkdir('./{}/'.format(dataset))
else:
shutil.rmtree('{}'.format(dataset))
os.mkdir('./{}/'.format(dataset))
# 将图片，标签，最不可能分类的标签，目标标签存入clean_datasets/CIFAR10/CIFAR10_inputs.npy等等或者MNIST
np.save('./{}/{}_inputs.npy'.format(dataset, dataset), candidate_images)
np.save('./{}/{}_labels.npy'.format(dataset, dataset), candidate_labels)
np.save('./{}/{}_llc.npy'.format(dataset, dataset), candidates_llc)
np.save('./{}/{}_targets.npy'.format(dataset, dataset), candidate_targets)
def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_index
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# dataset：数据集名称转化为大写形式MNIST/CIFAR10
dataset = args.dataset.upper()
assert dataset == 'MNIST' or dataset == 'CIFAR10'
# batch_size：每个分组的大小为1000
batch_size = 1000
# 获取MNIST/CIFAR10的模型，测试集
model_location = '{}/{}/model/{}_raw.pt'.format('../data', dataset,
dataset)
if dataset == 'MNIST':
raw_model = MNIST_CNN().to(device)
test_loader = get_mnist_test_loader(dir_name='../data/MNIST/',
batch_size=batch_size)
else:
raw_model = ResNet18().to(device)
test_loader = get_cifar10_test_loader(dir_name='../data/CIFAR10/',
batch_size=batch_size)
# 加载MNIST/CIFAR10的模型
raw_model.load(path=model_location, device=device)
# defense_name：防御名称为NRC
defense_name = 'NRC'
# 将参数传入NRC防御中
nrc = NRCDefense(model=raw_model,
defense_name=defense_name,
dataset=dataset,
device=device,
num_points=args.num_points)
# 如果要进行最优半径的搜索
if args.search:
# get the validation dataset (10% with the training dataset)
print('start to search the radius r using validation dataset ...')
# 获取MNIST/CIFAR10的验证集
if dataset == 'MNIST':
_, valid_loader = get_mnist_train_validate_loader(
dir_name='../data/MNIST/',
batch_size=batch_size,
valid_size=0.02,
shuffle=True)
else:
_, valid_loader = get_cifar10_train_validate_loader(
dir_name='../data/CIFAR10/',
batch_size=batch_size,
valid_size=0.02,
shuffle=True)
# radius：通过验证集得到最优的半径值
radius = nrc.search_best_radius(validation_loader=valid_loader,
radius_min=args.radius_min,
radius_max=args.radius_max,
radius_step=args.radius_step)
# 否则半径值为默认的0.01
else:
radius = round(args.radius, 2)
print(
'######\nthe radius for NRC is set or searched as: {}\n######'.format(
radius))
# 计算NRC模型在测试集上的分类精度
print(
'\nStart to calculate the accuracy of region-based classification defense on testing dataset'
)
raw_model.eval()
total = 0.0
correct = 0.0
with torch.no_grad():
for images, labels in test_loader:
nrc_labels = nrc.region_based_classification(samples=images,
radius=radius,
mean=args.mean,
std=args.std)
nrc_labels = torch.from_numpy(nrc_labels)
total += labels.size(0)
correct += (nrc_labels == labels).sum().item()
ratio = correct / total
print(
'\nTest accuracy of the {} model on the testing dataset: {:.1f}/{:.1f} = {:.2f}%\n'
.format(raw_model.model_name, correct, total, ratio * 100))