from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
import numpy as np
import matplotlib.pyplot as plt

# NOTE: This is a hack to get around "User-agent" limitations when downloading MNIST datasets
#       see, https://github.com/pytorch/vision/issues/3497 for more information
# from six.moves import urllib
# opener = urllib.request.build_opener()
# opener.addheaders = [('User-agent', 'Mozilla/5.0')]
# urllib.request.install_opener(opener)

epsilons = [0, .05, .1, .15, .2, .25, .3]
pretrained_model = "./data/lenet_mnist_model.pth"
use_cuda=True

# LeNet 模型定义
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, kernel_size=5) # 卷积层
        self.conv2 = nn.Conv2d(10, 20, kernel_size=5) # 卷积层
        self.conv2_drop = nn.Dropout2d() # dropout 层
        self.fc1 = nn.Linear(320, 50)
        self.fc2 = nn.Linear(50, 10)

    def forward(self, x): 
        x = F.relu(F.max_pool2d(self.conv1(x), 2)) # 卷积层
        x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) # 卷积层
        x = x.view(-1, 320) # 展平
        x = F.relu(self.fc1(x)) # 全连接层
        x = F.dropout(x, training=self.training) # dropout 层
        x = self.fc2(x) # 全连接层
        return F.log_softmax(x, dim=1) # 输出


# MNIST 测试数据集和数据加载器声明
test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('./data', 
                   train=False, # 不使用训练集
                   download=True,  # 下载数据集
                   transform=transforms.Compose([
                        transforms.ToTensor(), 
                    ])),
    batch_size=1, shuffle=True) # 批大小为 1，打乱数据

# 自动选择 GPU 或 CPU
print("CUDA Available: ",torch.cuda.is_available())
device = torch.device("cuda" if (use_cuda and torch.cuda.is_available()) else "cpu")

# 初始化网络
model = Net().to(device)

# 加载预训练模型
model.load_state_dict(torch.load(pretrained_model, map_location='cpu'))

# 将模型设置为评估模式。在本例中，这是针对 Dropout layers 的
model.eval()

# FGSM attack code
def fgsm_attack(image, epsilon, data_grad):
    # 收集数据梯度的元素符号
    sign_data_grad = data_grad.sign()
    # 通过调整输入图像的每个像素来创建扰动图像
    perturbed_image = image + epsilon*sign_data_grad
    # 添加剪切以保持 [0，1] 范围
    perturbed_image = torch.clamp(perturbed_image, 0, 1)
    # 返回被扰动的图像
    return perturbed_image

def test( model, device, test_loader, epsilon):

    # 正确率计数器
    correct = 0
    adv_examples = []

    # 循环访问测试集中的所有示例
    for data, target in test_loader:

        # 将数据和标签发送到 device
        data, target = data.to(device), target.to(device)

        # 设置 tensor 的 requires_grad 属性。这对攻击很重要
        data.requires_grad = True

        # 通过模型向前传递数据
        output = model(data)
        init_pred = output.max(1, keepdim=True)[1] # 获取最大对数概率（log-probability）的索引 

        # 如果最初的预测是错误的，直接跳过
        if init_pred.item() != target.item():
            continue

        # 计算 loss
        loss = F.nll_loss(output, target)

        # Zero all existing gradients
        model.zero_grad()

        # Calculate gradients of model in backward pass
        loss.backward()

        # 收集数据梯度
        data_grad = data.grad.data

        # Call FGSM Attack
        perturbed_data = fgsm_attack(data, epsilon, data_grad)

        # 对扰动图像重新分类
        output = model(perturbed_data)

        # 检查是否成功
        final_pred = output.max(1, keepdim=True)[1] # 获取最大对数概率的索引
        if final_pred.item() == target.item():
            correct += 1
            # 保存 0 epsilon 示例的特殊情况
            if (epsilon == 0) and (len(adv_examples) < 5):
                adv_ex = perturbed_data.squeeze().detach().cpu().numpy() 
                adv_examples.append( (init_pred.item(), final_pred.item(), adv_ex) )
        else:
            # 保存一些攻击成功示例以供以后可视化
            if len(adv_examples) < 5:
                adv_ex = perturbed_data.squeeze().detach().cpu().numpy()
                adv_examples.append( (init_pred.item(), final_pred.item(), adv_ex) )

    # 计算此 epsilon 的最终精度
    final_acc = correct/float(len(test_loader))
    print("Epsilon: {}\tTest Accuracy = {} / {} = {}".format(epsilon, correct, len(test_loader), final_acc))

    # 返回准确性和对抗性示例
    return final_acc, adv_examples

accuracies = []
examples = []

# 对每个ε运行测试
for eps in epsilons:
    acc, ex = test(model, device, test_loader, eps)
    accuracies.append(acc)
    examples.append(ex)

plt.figure(figsize=(5,5))
plt.plot(epsilons, accuracies, "*-")
plt.yticks(np.arange(0, 1.1, step=0.1))
plt.xticks(np.arange(0, .35, step=0.05))
plt.title("Accuracy vs Epsilon")
plt.xlabel("Epsilon")
plt.ylabel("Accuracy")
plt.show()

# 绘制每个 epsilon 处的对抗性样本的几个示例
cnt = 0
plt.figure(figsize=(8,10)) 
for i in range(len(epsilons)):
    for j in range(len(examples[i])):
        cnt += 1
        plt.subplot(len(epsilons),len(examples[0]),cnt)
        plt.xticks([], [])
        plt.yticks([], [])
        if j == 0:
            plt.ylabel("Eps: {}".format(epsilons[i]), fontsize=14)
        orig,adv,ex = examples[i][j]
        plt.title("{} -> {}".format(orig, adv),color=("green" if orig==adv else "red"))
        plt.imshow(ex, cmap="gray")
plt.tight_layout()
plt.show()