def perturb(self, X_nat, y):
"""
Given examples (X_nat, y), returns adversarial
examples within epsilon of X_nat in l_infinity norm.
"""
if self.rand:
X = X_nat + np.random.uniform(-self.epsilon, self.epsilon,
X_nat.shape).astype('float32')
else:
X = np.copy(X_nat)
for i in range(self.k):
X_var = base.to_var(torch.from_numpy(X), requires_grad=True)
y_var = base.to_var(torch.LongTensor(y))
scores = self.model(X_var)
loss = self.loss_fn(scores, y_var)
loss.backward()
grad = X_var.grad.data.cpu().numpy()
X += self.a * np.sign(grad)
X = np.clip(X, X_nat - self.epsilon, X_nat + self.epsilon)
X = np.clip(X, 0, 1) # ensure valid pixel range
X=torch.from_numpy(X)
return X
def perturb(self, X_nat, y, epsilons=None):
"""
Given examples (X_nat, y), returns their adversarial
counterparts with an attack length of epsilon.
"""
# Providing epsilons in batch
if epsilons is not None:
self.epsilon = epsilons
X = np.copy(X_nat)
X_var = base.to_var(torch.from_numpy(X), requires_grad=True)
y_var = base.to_var(torch.LongTensor(y))
scores = self.model(X_var)
loss = self.loss_fn(scores, y_var)
loss.backward()
grad_sign = X_var.grad.data.cpu().sign().numpy()
X += self.epsilon * grad_sign
X = np.clip(X, 0, 1)
X = torch.from_numpy(X)
return X
'''
训练模型
'''
net.train() # 训练模式
criterion = nn.CrossEntropyLoss() # 损失函数
optimizer = torch.optim.RMSprop(net.parameters(), lr=param['learning_rate'],
weight_decay=param['weight_decay']) # 优化器，具体怎么优化，学习率、正则化等等
'''对抗样本攻击'''
#adversary = Adversary.FGSM.FGSM(net, param['epsilon']) # 攻击方法
adversary = Adversary.LinfPGD.LinfPGDAttack(net, param['epsilon'])
for epoch in range(param['num_epochs']):
print('Starting epoch %d / %d' % (epoch + 1, param['num_epochs']))
for t, (x, y) in enumerate(loader_train):
x_var, y_var = base.to_var(x), base.to_var(y.long()) # 转换格式
X_adv = adversary.perturb(x.numpy(), y) # 加入扰动，生成对抗样本
loss = (criterion(net(x_var), y_var) + criterion(net(X_adv.cuda()), y_var))/2 # 计算损失
optimizer.zero_grad() # 把上一轮的梯度清零
loss.backward() # 反向传播求导
optimizer.step() # 优化参数
if (t + 1) % 10 == 0: # 每训练10批数据看一下loss
print('t = %d, loss = %.8f' % (t + 1, loss.item()))
if (t + 1) % 100 == 0:
num_correct, num_samples, acc = Optimizer.test(net, loader_test)
print('[train] t = %d, right predict:(%d/%d) ,pre test_acc=%.4f%%' % (t + 1, num_correct, num_samples, acc))
torch.save(net.state_dict(), modelpath) # 保存模型到文件