def adv_train_process(self,delay=0.5):
"Adversarial training, returns pertubed mini batch"
"delay: parameter to decide how many epochs should be used as adv training"
if self.opts['Adversary'] == 'WT':
adv_train_data_path = '../data/WalkieTalkie/defended_csv/adv_train_WT.csv'
else:
adv_train_data_path = self.data_path + self.classifier_type + '/' + self.opts['adv_train_data_path']
print('adv train data path: ',adv_train_data_path)
if self.mode in ['wf','wf_ow','wf_kf']:
"load data"
if self.mode != 'wf_kf':
train_path = self.data_path + self.opts['train_data_path']
train_data = utils_wf.load_data_main(train_path,self.opts['batch_size'],shuffle=True)
# test_data = utils_wf.load_data_main(self.data_path + self.opts['test_data_path'],self.opts['batch_size'])
else:
# benign_path = '../data/wf/cross_val/'
# train_path = benign_path + self.opts['train_data_path']
train_path = '../data/wf/train_NoDef_burst.csv'
train_data = utils_wf.load_data_main(train_path,self.opts['batch_size'], shuffle=True)
# test_data = utils_wf.load_data_main(benign_path + self.opts['test_data_path'],self.opts['batch_size'])
print('train data path: ',train_path)
adv_train_data = utils_wf.load_data_main(adv_train_data_path,self.opts['batch_size'],shuffle=True)
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'], self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
target_model.train()
elif self.classifier_type == 'lstm':
if self.mode == 'wf_ow':
params = utils_wf.params_lstm_ow(self.opts['num_class'], self.opts['input_size'],self.opts['batch_size'])
else:
params = utils_wf.params_lstm(self.opts['num_class'], self.opts['input_size'],self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
target_model.train()
elif self.mode == 'shs':
"load data"
train_data = utils_shs.load_data_main(self.data_path + self.opts['train_data_path'],self.opts['batch_size'])
test_data = utils_shs.load_data_main(self.data_path + self.opts['test_data_path'],self.opts['batch_size'])
adv_train_data = utils_shs.load_data_main(adv_train_data_path, self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_shs.params_cnn(self.opts['num_class'], self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
target_model.train()
elif self.classifier_type == 'lstm':
params = utils_shs.params_lstm(self.opts['num_class'], self.opts['input_size'], self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
target_model.train()
else:
print('mode not in ["wf","shs"], system will exit.')
sys.exit()
loss_criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(target_model.parameters(), lr=1e-3)
"start training"
steps = 0
flag = False
start_time = datetime.now()
# print('adversary working dir {}'.format(adv_train_data_path))
for epoch in range(self.opts['epochs']):
print('Starting epoch %d / %d' % (epoch + 1, self.opts['epochs']))
if flag:
print('{} based adversarial training...'.format(self.opts['Adversary']))
for (x,y),(x_adv,y_adv) in zip(train_data,adv_train_data):
steps += 1
optimizer.zero_grad()
x,y = x.to(self.device), y.to(self.device)
outputs = target_model(x)
loss = loss_criterion(outputs,y)
"adversarial training"
if epoch + 1 >= int((1-delay)*self.opts['epochs']):
x_adv, y_adv = x_adv.to(self.device), y_adv.to(self.device)
flag = True
loss_adv = loss_criterion(target_model(x_adv),y_adv)
loss = (loss + loss_adv) / 2
"print results every 100 steps"
if steps % 100 == 0:
end_time = datetime.now()
time_diff = (end_time - start_time).seconds
time_usage = '{:3}m{:3}s'.format(int(time_diff / 60), time_diff % 60)
msg = "Step {:5}, Loss:{:6.2f}, Time usage:{:9}."
print(msg.format(steps, loss, time_usage))
loss.backward()
optimizer.step()
if epoch!=0 and epoch % 10 == 0 or epoch == self.opts['epochs']:
if self.mode != 'wf_kf':
output_name = '/adv_target_model_%s.pth' % self.opts['Adversary']
else:
output_name = '/adv_target_model_%s_%d.pth' % (self.opts['Adversary'], self.opts['id'])
output_path = self.model_path + output_name
torch.save(target_model.state_dict(), output_path)
"test trianed model"
def test_peformance(self, threshold_val):
"load data and target model"
if self.mode == 'wf_ow':
"load data"
if self.opts['test_data_type'] == 'NoDef':
test_data_Mon = utils_wf.load_data_main(
self.opts['test_data_Mon_path'], self.opts['batch_size'])
test_data_UnMon = utils_wf.load_data_main(
self.opts['test_data_UnMon_path'], self.opts['batch_size'])
elif self.opts['test_data_type'] == 'Def':
test_data_Mon = utils_wf.load_data_main(
self.opts['adv_test_data_Mon_path'],
self.opts['batch_size'])
test_data_UnMon = utils_wf.load_data_main(
self.opts['adv_test_data_UnMon_path'],
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
params = utils_wf.params_lstm_ow_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
else:
print('mode not in ["wf_ow"], system will exit.')
sys.exit()
if self.model_type == 'adv_target_model':
model_name = self.model_path + '/adv_target_model_' + self.opts[
'Adversary'] + '.pth'
elif self.model_type == 'target_model':
model_name = self.model_path + '/target_model.pth'
else:
print(
'target model type not in ["target_model","adv_target_model"], system will exit.'
)
sys.exit()
target_model.load_state_dict(
torch.load(model_name, map_location=self.device))
target_model.eval()
print('target model %s' % model_name)
print('test data path unmonitored %s ' %
self.opts['adv_test_data_UnMon_path'])
print('test data path monitored %s ' %
self.opts['adv_test_data_Mon_path'])
# ==============================================================
"testing process..."
TP = 0
FP = 0
TN = 0
FN = 0
"obtain monitored and Unmonitored lables"
_, y_test_Mon = utils_wf.load_csv_data(self.opts['test_data_Mon_path'])
_, y_test_UnMon = utils_wf.load_csv_data(
self.opts['test_data_UnMon_path'])
monitored_labels = torch.Tensor(y_test_Mon).long().to(self.device)
unmonitored_labels = torch.Tensor(y_test_UnMon).long().to(self.device)
# ==============================================================
# Test with Monitored testing instances
for i, data in enumerate(test_data_Mon, 0):
test_x, test_y = data
test_x, test_y = test_x.to(self.device), test_y.to(self.device)
max_probs, pred_labels = torch.max(
torch.softmax(target_model(test_x), 1), 1)
for j, pred_label in enumerate(pred_labels):
if pred_label in monitored_labels: # predited as Monitored
if max_probs[
j] >= threshold_val: # probability greater than the threshold
TP += 1
else: # predicted as unmonitored and true lable is Monitored
FN += 1
elif pred_label in unmonitored_labels: # predicted as unmonitored and true lable is monitored
FN += 1
# ==============================================================
# Test with Unmonitored testing instances
for i, data in enumerate(test_data_UnMon, 0):
test_x, test_y = data
test_x, test_y = test_x.to(self.device), test_y.to(self.device)
max_probs, pred_labels = torch.max(
torch.softmax(target_model(test_x), 1), 1)
for j, pred_label in enumerate(pred_labels):
if pred_label in unmonitored_labels: # predited as unmonitored and true label is unmonitored
TN += 1
elif pred_label in monitored_labels: # predicted as Monitored
if max_probs[
j] >= threshold_val: # predicted as monitored and true label is unmonitored
FP += 1
else:
TN += 1
"print result"
print("TP : ", TP)
print("FP : ", FP)
print("TN : ", TN)
print("FN : ", FN)
print("Total : ", TP + FP + TN + FN)
TPR = float(TP) / (TP + FN)
print("TPR : ", TPR)
FPR = float(FP) / (FP + TN)
print("FPR : ", FPR)
Precision = float(TP) / (TP + FP)
print("Precision : ", Precision)
Recall = float(TP) / (TP + FN)
print("Recall : ", Recall)
print("\n")
self.log_file.writelines(
"%.6f,%d,%d,%d,%d,%.6f,%.6f,%.6f,%.6f\n" %
(threshold_val, TP, FP, TN, FN, TPR, FPR, Precision, Recall))
def train(self,train_dataloader):
"load target model"
if self.mode == 'wf' or 'wf_ow' or 'wf_kf':
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'], self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
"eval: lstm can't work on eval model, so set dropout as 0 to make train model as eval model"
if self.mode == 'wf_ow':
params = utils_wf.params_lstm_ow_eval(self.opts['num_class'], self.opts['input_size'],self.opts['batch_size'])
else:
params = utils_wf.params_lstm_eval(self.opts['num_class'], self.opts['input_size'], self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_shs.params_cnn(self.opts['num_class'], self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
elif self.classifier_type == 'lstm':
"eval: lstm can't work on eval model, so set dropout as 0 to make train model as eval model"
params = utils_shs.params_lstm_eval(self.opts['num_class'], self.opts['input_size'], self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
else:
print('mode not in ["wf","shs","wf_ow","wf_kf"], system will exit.')
sys.exit()
if self.mode == 'wf_kf':
model_name = '/target_model_%d.pth' % self.opts['id']
else:
model_name = '/target_model.pth'
model_path = self.model_path + model_name
target_model.load_state_dict(torch.load(model_path, map_location=self.device))
"set equal_eval mode of train instead eval for lstm"
if self.classifier_type == 'lstm':
target_model = self.model_reset(target_model)
target_model.train()
elif self.classifier_type == 'cnn':
target_model.eval()
target_model.to(self.device)
for epoch in range(1, self.opts['epochs'] + 1):
if epoch == 50:
self.optimizer_G = torch.optim.Adam(self.generator.parameters(),
lr=0.0001)
self.optimizer_D = torch.optim.Adam(self.discriminator.parameters(),
lr=0.0001)
if epoch == 80:
self.optimizer_G = torch.optim.Adam(self.generator.parameters(),
lr=0.00001)
self.optimizer_D = torch.optim.Adam(self.discriminator.parameters(),
lr=0.00001)
loss_D_sum = 0
loss_G_fake_sum = 0
loss_perturb_sum = 0
loss_adv_sum = 0
for i, data in enumerate(train_dataloader, start=0):
train_x, train_y = data
train_x, train_y = train_x.to(self.device), train_y.to(self.device)
loss_D_batch, loss_G_fake_batch, loss_perturb_batch, loss_adv_batch = \
self.train_batch(train_x, train_y,target_model)
loss_D_sum += loss_D_batch
loss_G_fake_sum += loss_G_fake_batch
loss_perturb_sum += loss_perturb_batch
loss_adv_sum += loss_adv_batch
"print statistics"
num_batch = len(train_dataloader)
print("epoch %d:\nloss_D: %.3f, loss_G_fake: %.3f,\
\nloss_perturb: %.3f, loss_adv: %.3f, \n" %
(epoch, loss_D_sum / num_batch, loss_G_fake_sum / num_batch,
loss_perturb_sum / num_batch, loss_adv_sum / num_batch))
"save generator"
if epoch != 0 and epoch % 10 == 0:
if self.mode == 'wf_kf':
out_model_name = '/adv_generator_%d.pth' % self.opts['id']
else:
out_model_name = '/adv_generator.pth'
torch.save(self.generator.state_dict(), self.model_path + out_model_name)
def test_model(self):
"define test data type/path"
# classifier cross validation:
if self.opts['cross_validation']:
if self.classifier_type == 'lstm':
adv_test_data_path = '../data/' + self.mode + '/cnn/' + self.opts[
'adv_test_data_path']
elif self.classifier_type == 'cnn':
adv_test_data_path = '../data/' + self.mode + '/lstm/' + self.opts[
'adv_test_data_path']
else:
print(
'classifier type should in [lstm,cnn]. System will exit!')
sys.exit()
# test on Walkie-Talkie defended data
elif self.opts['Adversary'] == 'WT':
adv_test_data_path = '../data/WalkieTalkie/defended_csv/adv_test_WT.csv'
else:
adv_test_data_path = self.data_path + '/' + self.opts[
'adv_test_data_path']
print('adv test data path: ', adv_test_data_path)
"load data and model"
if self.mode == 'wf' or 'wf_ow' or 'wf_kf':
"load data"
if self.mode == 'wf_kf':
test_path = '../data/wf/cross_val/' + self.opts[
'test_data_path']
test_data = utils_wf.load_data_main(test_path,
self.opts['batch_size'])
else:
test_path = '../data/wf/' + self.opts['test_data_path']
test_data = utils_wf.load_data_main(test_path,
self.opts['batch_size'])
adv_test_data = utils_wf.load_data_main(adv_test_data_path,
self.opts['batch_size'])
print('test data path: ', test_path)
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
if self.mode == 'wf_ow':
params = utils_wf.params_lstm_ow_eval(
self.opts['num_class'], self.opts['input_size'],
self.opts['batch_size'])
else:
params = utils_wf.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load data"
test_data = utils_shs.load_data_main(
'../data/' + self.mode + '/' + self.opts['test_data_path'],
self.opts['batch_size'])
adv_test_data = utils_shs.load_data_main(adv_test_data_path,
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_shs.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
elif self.classifier_type == 'lstm':
params = utils_shs.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
else:
print('mode not in ["wf","shs"], system will exit.')
sys.exit()
if self.model_type == 'adv_target_model':
if self.mode == 'wf_kf':
model_name = self.model_path + '/adv_target_model_%s_%d.pth' % (
self.opts['Adversary'], self.opts['id'])
else:
model_name = self.model_path + '/adv_target_model_' + self.opts[
'Adversary'] + '.pth'
elif self.model_type == 'target_model':
if self.mode == 'wf_kf':
model_name = self.model_path + '/target_model_%d.pth' % self.opts[
'id']
else:
model_name = self.model_path + '/target_model.pth'
else:
print(
'target model type not in ["target_model","adv_target_model"], system will exit.'
)
sys.exit()
print('model path: ', model_name)
target_model.load_state_dict(
torch.load(model_name, map_location=self.device))
target_model.eval()
"test on adversarial examples"
correct_adv_x = 0
correct_x = 0
total_case = 0
for (x, y), (adv_x, adv_y) in zip(test_data, adv_test_data):
x, y = x.to(self.device), y.to(self.device)
adv_x, adv_y = adv_x.to(self.device), adv_y.to(self.device)
"prediction on original input x"
pred_x = target_model(x)
_, pred_x = torch.max(pred_x, 1)
correct_x += (pred_x == y).sum()
"predition on adv_x"
pred_adv_x = target_model(adv_x)
_, pred_adv_x = torch.max(pred_adv_x, 1)
correct_adv_x += (pred_adv_x == adv_y).sum()
total_case += len(y)
acc_x = float(correct_x.item()) / float(total_case)
acc_adv = float(correct_adv_x.item()) / float(total_case)
print('*' * 30)
print('"{}" with {} against {}.'.format(self.mode,
self.opts['Adversary'],
self.model_type))
print('correct test after attack is {}'.format(correct_adv_x.item()))
print('total test instances is {}'.format(total_case))
print(
'accuracy of test after {} attack : correct/total= {:.5f}'.format(
self.opts['Adversary'], acc_adv))
print('success rate of the attack is : {}'.format(1 - acc_adv))
print('accucary of the model without being attacked is {:.5f}'.format(
acc_x))
print('\n')
def ploting(self):
"load data and model"
if self.mode == 'wf':
"load data"
test_data = utils_wf.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
params = utils_wf.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load data"
test_data = utils_shs.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
params = utils_shs.params(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
else:
print('mode not in ["wf","shs"], system will exit.')
sys.exit()
model_name = self.model_path + '/target_model.pth'
target_model.load_state_dict(
torch.load(model_name, map_location=self.device))
"set equal_eval mode of train instead eval for lstm to aviod backward error"
if self.classifier_type == 'lstm':
target_model = self.model_reset(target_model)
target_model.train()
elif self.classifier_type == 'cnn':
target_model.eval()
target_model.to(self.device)
###############################################
for i, data in enumerate(test_data):
if i < self.opts['num_figs']:
for Adversary in ['FGSM', 'PGD', 'DeepFool', 'GAN']:
x, y = data
x, y = x.to(self.device), y.to(self.device)
"set adversary"
if self.mode == 'wf':
fgsm_epsilon = 0.1
pgd_a = 0.051
else:
fgsm_epsilon = 0.1
pgd_a = 0.01
if Adversary == 'GAN':
pretrained_generator_path = self.model_path + '/adv_generator.pth'
pretrained_G = models.Generator(
self.gen_input_nc, self.input_nc).to(self.device)
pretrained_G.load_state_dict(
torch.load(pretrained_generator_path,
map_location=self.device))
pretrained_G.eval()
elif Adversary == 'FGSM':
adversary = FGSM(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
epsilon=fgsm_epsilon)
elif Adversary == 'DeepFool':
adversary = DeepFool(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
num_classes=5)
elif Adversary == 'PGD':
adversary = LinfPGDAttack(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
k=5,
a=pgd_a,
random_start=False)
"generate adversarial examples"
if Adversary == 'GAN':
pert = pretrained_G(x)
"cal adv_x given different mode wf/shs"
adv_x = utils_gan.get_advX_gan(
x,
pert,
self.mode,
pert_box=self.opts['pert_box'],
x_box_min=self.opts['x_box_min'],
x_box_max=self.opts['x_box_max'],
alpha=self.opts['alpha'])
elif Adversary in ['FGSM', 'PGD', 'DeepFool']:
_, y_pred = torch.max(target_model(x), 1)
"cal adv_x given different mode wf/shs. the mode of adversary set before"
adversary.model = target_model
"use predicted label to prevent label leaking"
_, adv_x = adversary.perturbation(
x, y_pred, self.opts['alpha'])
else:
print('No Adversary found! System will exit.')
sys.exit()
if self.mode == 'shs':
adv_x = (adv_x.data.cpu().numpy().squeeze() *
1500).round()
elif self.mode == 'wf':
"if the data use L2 normalized, then need to inverse it back"
# normalization = utils_wf.normalizer(x)
# adv_x = normalization.inverse_Normalizer(adv_x)
adv_x = adv_x.data.cpu().numpy().squeeze()
adv_x = adv_x.squeeze()
else:
print('mode should in ["wf","shs"], system will exit.')
sys.exit()
x_np = x.data.cpu().numpy().squeeze()
pert = adv_x - x_np
# traffics.append(np.squeeze(x.data.cpu().numpy().squeeze()))
# adv_traffics.append(adv_x)
# noise.append(pert)
"plot"
utils_shs.single_traffic_plot(
self.classifier_type + '_' + str(i), x_np, adv_x,
Adversary)
utils_shs.noise_plot(self.classifier_type + '_' + str(i),
pert, Adversary)
def generate(self):
"generate adv_x given x, append with its original label y instead with y_pert "
"load data and target model"
if self.mode in ['wf', 'wf_ow', 'wf_kf', 'wf_sg']:
"load data"
if self.mode != 'wf_sg':
train_data = utils_wf.load_data_main(
self.opts['train_data_path'], self.opts['batch_size'])
test_data = utils_wf.load_data_main(
self.opts['test_data_path'], self.opts['batch_size'])
else:
input_data = utils_wf.load_data_main(
self.opts['input_data_path'], self.opts['batch_size'])
train_data, test_data = []
if self.mode == 'wf_ow':
test_data_UnMon = utils_wf.load_data_main(
self.opts['test_data_path_UnMon'], self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
if self.mode == 'wf_ow':
params = utils_wf.params_lstm_ow_eval(
self.opts['num_class'], self.opts['input_size'],
self.opts['batch_size'])
else:
params = utils_wf.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load data"
train_data = utils_shs.load_data_main(self.opts['train_data_path'],
self.opts['batch_size'])
test_data = utils_shs.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_shs.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
elif self.classifier_type == 'lstm':
params = utils_shs.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
else:
print(
'mode not in ["wf","shs","wf_ow","wf_kf"], system will exit.')
sys.exit()
# load trained model
if self.mode == 'wf_kf':
model_name = '/target_model_%d.pth' % self.opts['id']
else:
model_name = '/target_model.pth'
model_path = self.model_path + model_name
target_model.load_state_dict(
torch.load(model_path, map_location=self.device))
"set adversary"
Adversary = self.opts['Adversary']
if self.mode == 'wf' or 'wf_ow' or 'wf_kf':
fgsm_epsilon = 0.1
pgd_a = 0.051 # if data un-normalized
# pgd_a = 0.01 # if data normalized
else:
fgsm_epsilon = 0.1
pgd_a = 0.01
if Adversary == 'GAN':
if self.mode == 'wf_kf':
g_name = '/adv_generator_%d.pth' % self.opts['id']
else:
g_name = '/adv_generator.pth'
pretrained_generator_path = self.model_path + g_name
adversary = models.Generator(self.gen_input_nc,
self.input_nc).to(self.device)
adversary.load_state_dict(
torch.load(pretrained_generator_path,
map_location=self.device))
adversary.eval()
elif Adversary == 'FGSM':
adversary = FGSM(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
epsilon=fgsm_epsilon)
elif Adversary == 'DeepFool':
adversary = DeepFool(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
num_classes=5)
elif Adversary == 'PGD':
adversary = LinfPGDAttack(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
k=5,
a=pgd_a,
random_start=False)
"produce adv_x given differen input data"
# test data Monitored
data_type = 'test'
self.get_adv_x(test_data, Adversary, adversary, target_model,
data_type)
# test data Unmonitored in open world setting
if self.mode == 'wf_ow':
data_type = 'test_UnMon'
self.get_adv_x(test_data_UnMon, Adversary, adversary, target_model,
data_type)
#train data
data_type = 'train'
self.get_adv_x(train_data, Adversary, adversary, target_model,
data_type)
def train_model(self):
if self.mode in ['wf', 'detect', 'wf_ow', 'wf_kf']:
"load data"
train_data = utils_wf.load_data_main(self.opts['train_data_path'],
self.opts['batch_size'],
shuffle=True)
test_data = utils_wf.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'],
shuffle=True)
"load target model structure"
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
target_model.train()
elif self.mode == 'shs':
"load data"
train_data = utils_shs.load_data_main(self.opts['train_data_path'],
self.opts['batch_size'])
test_data = utils_shs.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
params = utils_shs.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
target_model.train()
else:
print(
'mode not in ["wf","shs","wf_ow","detect","wf-kf"], system will exit.'
)
sys.exit()
"train process"
optimizer = torch.optim.Adam(target_model.parameters(), lr=0.001)
for epoch in range(self.opts['epochs']):
loss_epoch = 0
for i, data in enumerate(train_data, 0):
train_x, train_y = data
train_x, train_y = train_x.to(self.device), train_y.to(
self.device)
optimizer.zero_grad()
logits_model = target_model(train_x)
loss_model = F.cross_entropy(logits_model, train_y)
loss_epoch += loss_model
loss_model.backward()
optimizer.step()
if i % 100 == 0:
_, predicted = torch.max(logits_model, 1)
correct = int(sum(predicted == train_y))
accuracy = correct / len(train_y)
msg = 'Epoch {:5}, Step {:5}, Loss: {:6.2f}, Accuracy:{:8.2%}.'
print(msg.format(epoch, i, loss_model, accuracy))
"save model every 10 epochs"
if self.mode == 'wf_kf':
model_name = '/target_model_%d.pth' % self.opts['id']
else:
model_name = '/target_model.pth'
if epoch != 0 and epoch % 10 == 0:
targeted_model_path = self.model_path + model_name
torch.save(target_model.state_dict(), targeted_model_path)
"test target model"
target_model.eval()
num_correct = 0
total_instances = 0
for i, data in enumerate(test_data, 0):
test_x, test_y = data
test_x, test_y = test_x.to(self.device), test_y.to(self.device)
pred_lab = torch.argmax(target_model(test_x), 1)
num_correct += torch.sum(pred_lab == test_y, 0)
total_instances += len(test_y)
print('accuracy of target model against test dataset: %f\n' %
(num_correct.item() / total_instances))
def test_peformance(self):
"load data and target model"
if self.mode == 'wf' or 'detect' or 'wf_ow' or 'wf_tf':
"load data"
test_data = utils_wf.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
if self.mode == 'wf_ow':
params = utils_wf.params_lstm_ow(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
elif self.mode == 'wf_kf':
params = utils_wf.params_lstm(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
else:
params = utils_wf.params_lstm(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load data"
test_data = utils_shs.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
params = utils_shs.params(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
else:
print('mode not in ["wf","shs"], system will exit.')
sys.exit()
if self.mode == 'wf_kf':
model_name = '/target_model_%d.pth' % self.opts['id']
else:
model_name = '/target_model.pth'
model_path = self.model_path + model_name
print(model_path)
target_model.load_state_dict(
torch.load(model_path, map_location=self.device))
target_model.eval()
"testing process..."
num_correct = 0
total_case = 0
y_test = []
y_pred = []
for i, data in enumerate(test_data, 0):
test_x, test_y = data
test_x, test_y = test_x.to(self.device), test_y.to(self.device)
"add softmax after fc of model to normalize output as positive values and sum =1"
pred_lab = torch.argmax(torch.softmax(target_model(test_x), 1), 1)
# pred_lab = torch.argmax(target_model(test_x), 1)
num_correct += torch.sum(pred_lab == test_y, 0)
total_case += len(test_y)
"save result"
y_test += (test_y.cpu().numpy().tolist())
y_pred += (pred_lab.cpu().numpy().tolist())
# print('accuracy in testing set: %f\n' % (num_correct.item() / total_case))
# print(classification_report(y_test, y_pred))
print('confusion matrix is {}'.format(confusion_matrix(y_test,
y_pred)))
print('accuracy is {}'.format(metrics.accuracy_score(y_test, y_pred)))
def test_model(self):
"load data and model"
if self.mode == 'wf':
"load data"
test_data = utils_wf.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
if self.classifier_type == 'cnn':
params = utils_wf.params_cnn(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_norm(params).to(self.device)
elif self.classifier_type == 'lstm':
params = utils_wf.params_lstm_eval(self.opts['num_class'],
self.opts['input_size'],
self.opts['batch_size'])
target_model = models.lstm(params).to(self.device)
elif self.mode == 'shs':
"load data"
test_data = utils_shs.load_data_main(self.opts['test_data_path'],
self.opts['batch_size'])
"load target model structure"
params = utils_shs.params(self.opts['num_class'],
self.opts['input_size'])
target_model = models.cnn_noNorm(params).to(self.device)
else:
print('mode not in ["wf","shs"], system will exit.')
sys.exit()
if self.target_model_type == 'adv_target_model':
model_name = self.model_path + '/adv_target_model_' + self.opts[
'Adversary'] + '.pth'
elif self.target_model_type == 'target_model':
model_name = self.model_path + '/target_model.pth'
else:
print(
'target model type not in ["target_model","adv_target_model"], system will exit.'
)
sys.exit()
target_model.load_state_dict(
torch.load(model_name, map_location=self.device))
"set equal_eval mode of train instead eval for lstm"
if self.classifier_type == 'lstm':
target_model = self.model_reset(target_model)
target_model.train()
elif self.classifier_type == 'cnn':
target_model.eval()
target_model.to(self.device)
"set adversary"
Adversary = self.opts['Adversary']
if self.mode == 'wf':
fgsm_epsilon = 0.1
pgd_a = 0.051
elif self.mode == 'shs':
fgsm_epsilon = 0.1
pgd_a = 0.01
if Adversary == 'GAN':
pretrained_generator_path = self.model_path + '/adv_generator.pth'
pretrained_G = models.Generator(self.gen_input_nc,
self.input_nc).to(self.device)
pretrained_G.load_state_dict(
torch.load(pretrained_generator_path,
map_location=self.device))
pretrained_G.eval()
elif Adversary == 'FGSM':
adversary = FGSM(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
epsilon=fgsm_epsilon)
elif Adversary == 'DeepFool':
adversary = DeepFool(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
num_classes=5)
elif Adversary == 'PGD':
adversary = LinfPGDAttack(self.mode,
self.x_box_min,
self.x_box_max,
self.pert_box,
k=5,
a=pgd_a,
random_start=False)
"test on adversarial examples"
num_correct = 0
correct_x = 0
total_case = 0
for i, data in enumerate(test_data, 0):
test_x, test_y = data
test_x, test_y = test_x.to(self.device), test_y.to(self.device)
"prediction on original input x"
pred_y = target_model(test_x)
_, pred_y = torch.max(pred_y, 1)
"prediction on adversarial x"
if Adversary in ['FGSM', 'DeepFool', 'PGD']:
adversary.model = target_model
"use predicted label to prevent label leaking"
adv_y, adv_x = adversary.perturbation(test_x, pred_y,
self.opts['alpha'])
# adv_y,adv_x = adversary.perturbation(test_x,test_y,self.opts['alpha'])
elif Adversary == 'GAN':
pert = pretrained_G(test_x)
adv_x = utils_gan.get_advX_gan(test_x, pert, self.mode,
self.pert_box, self.x_box_min,
self.x_box_max,
self.opts['alpha'])
adv_y = torch.argmax(target_model(adv_x.to(self.device)), 1)
num_correct += torch.sum(adv_y == test_y, 0)
correct_x += (pred_y == test_y).sum()
total_case += len(test_y)
acc = float(num_correct.item()) / float(total_case)
print('*' * 30)
print('"{}" with {} against {}.'.format(self.mode, Adversary,
self.target_model_type))
print('correct test after attack is {}'.format(num_correct.item()))
print('total test instances is {}'.format(total_case))
print(
'accuracy of test after {} attack : correct/total= {:.5f}'.format(
Adversary, acc))
print('success rate of the attack is : {}'.format(1 - acc))
print('accucary of the model without being attacked is {:.5f}'.format(
float(correct_x) / float(total_case)))
print('\n')