学习参考来自
使用CNN在MNIST上实现简单的攻击样本https://github.com/wmn7/ML_Practice/blob/master/2019_06_03/CNN_MNIST%E5%8F%AF%E8%A7%86%E5%8C%96.ipynb文章目录 在 MNIST 上实现简单的攻击样本1 训练一个数字分类网络2 控制输出的概率, 看输入是什么3 让正确的图片分类错误
在 MNIST 上实现简单的攻击样本
要看某个filter在检测什么，我们让通过其后的输出激活值较大。
想要一些攻击样本，有一个简单的想法就是让最后结果中是某一类的概率值变大，接着进行反向传播(固定住网络的参数)，去修改input
原理图：
1 训练一个数字分类网络模型基础数据、超参数的配置
import timeimport csv, osimport PIL.Imageimport numpy as npimport pandas as pdimport matplotlib.pyplot as pltimport cv2from cv2 import resizeimport torchimport torch.nn as nnimport torch.nn.functional as Ffrom torch.utils.data.sampler import SubsetRandomSamplerfrom torch.autograd import Variableimport copyimport torchvisionimport torchvision.transforms as transformsos.environ["CUDA_VISIBLE_DEVICES"] = "1"# --------------------# Device configuration# --------------------device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')# ----------------# Hyper-parameters# ----------------num_classes = 10num_epochs = 3batch_size = 100validation_split = 0.05 # 每次训练集中选出10%作为验证集learning_rate = 0.001# -------------# MNIST dataset# -------------train_dataset = torchvision.datasets.MNIST(root='./',train=True,transform=transforms.ToTensor(),download=True)test_dataset = torchvision.datasets.MNIST(root='./',train=False,transform=transforms.ToTensor())# -----------# Data loader# -----------test_len = len(test_dataset) # 计算测试集的个数 10000test_loader = torch.utils.data.DataLoader(dataset=test_dataset,batch_size=batch_size,shuffle=False)for (inputs, labels) in test_loader:print(inputs.size()) # [100, 1, 28, 28]print(labels.size()) # [100]break# ------------------# 下面切分validation# ------------------dataset_len = len(train_dataset) # 60000indices = list(range(dataset_len))# Randomly splitting indices:val_len = int(np.floor(validation_split * dataset_len)) # validation的长度validation_idx = np.random.choice(indices, size=val_len, replace=False) # validatiuon的indextrain_idx = list(set(indices) - set(validation_idx)) # train的index## Defining the samplers for each phase based on the random indices:train_sampler = SubsetRandomSampler(train_idx)validation_sampler = SubsetRandomSampler(validation_idx)train_loader = torch.utils.data.DataLoader(dataset=train_dataset,sampler=train_sampler,batch_size=batch_size)validation_loader = torch.utils.data.DataLoader(train_dataset,sampler=validation_sampler,batch_size=batch_size)train_dataloaders = {"train": train_loader, "val": validation_loader} # 使用字典的方式进行保存train_datalengths = {"train": len(train_idx), "val": val_len} # 保存train和validation的长度搭建简单的神经网络
# -------------------------------------------------------# Convolutional neural network (two convolutional layers)# -------------------------------------------------------class ConvNet(nn.Module):def __init__(self, num_classes=10):super(ConvNet, self).__init__()self.layer1 = nn.Sequential(nn.Conv2d(in_channels=1, out_channels=16, kernel_size=5, stride=1, padding=2),nn.BatchNorm2d(16),nn.ReLU(),nn.MaxPool2d(kernel_size=2, stride=2))self.layer2 = nn.Sequential(nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5, stride=1, padding=2),nn.BatchNorm2d(32),nn.ReLU(),nn.MaxPool2d(kernel_size=2, stride=2))self.fc = nn.Linear(7*7*32, num_classes)def forward(self, x):out = self.layer1(x)out = self.layer2(out)out = out.reshape(out.size(0), -1)out = self.fc(out)return out绘制损失和精度变化曲线
# -----------# draw losses# -----------def draw_loss_acc(train_list, validation_list, mode="loss"):plt.style.use("seaborn")# set interdata_len = len(train_list)x_ticks = np.arange(1, data_len + 1)plt.xticks(x_ticks)if mode == "Loss":plt.plot(x_ticks, train_list, label="Train Loss")plt.plot(x_ticks, validation_list, label="Validation Loss")plt.xlabel("Epoch")plt.ylabel("Loss")plt.legend()plt.savefig("Epoch_loss.jpg")elif mode == "Accuracy":plt.plot(x_ticks, train_list, label="Train AccuracyAccuracy")plt.plot(x_ticks, validation_list, label="Validation Accuracy")plt.xlabel("Epoch")plt.ylabel("Accuracy")plt.legend()plt.savefig("Epoch_Accuracy.jpg")定义训练模型的函数
# ---------------# Train the model# ---------------def train_model(model, criterion, optimizer, dataloaders, train_datalengths, scheduler=None, num_epochs=2):"""传入的参数分别是:1. model:定义的模型结构2. criterion:损失函数3. optimizer:优化器4. dataloaders:training dataset5. train_datalengths:train set和validation set的大小, 为了计算准确率6. scheduler:lr的更新策略7. num_epochs:训练的epochs"""since = time.time()# 保存最好一次的模型参数和最好的准确率best_model_wts = copy.deepcopy(model.state_dict())best_acc = 0.0train_loss = [] # 记录每一个epoch后的train的losstrain_acc = []validation_loss = []validation_acc = []for epoch in range(num_epochs):print('Epoch [{}/{}]'.format(epoch + 1, num_epochs))print('-' * 10)# Each epoch has a training and validation phasefor phase in ['train', 'val']:if phase == 'train':if scheduler != None:scheduler.step()model.train() # Set model to training modeelse:model.eval() # Set model to evaluate moderunning_loss = 0.0 # 这个是一个epoch积累一次running_corrects = 0 # 这个是一个epoch积累一次# Iterate over data.total_step = len(dataloaders[phase])for i, (inputs, labels) in enumerate(dataloaders[phase]):# inputs = inputs.reshape(-1, 28*28).to(device)inputs = inputs.to(device)labels = labels.to(device)# zero the parameter gradientsoptimizer.zero_grad()# forward# track history if only in trainwith torch.set_grad_enabled(phase == 'train'):outputs = model(inputs)_, preds = torch.max(outputs, 1) # 使用output(概率)得到预测loss = criterion(outputs, labels) # 使用output计算误差# backward + optimize only if in training phaseif phase == 'train':loss.backward()optimizer.step()# statisticsrunning_loss += loss.item() * inputs.size(0)running_corrects += torch.sum(preds == labels.data)if (i + 1) % 100 == 0:# 这里相当于是i*batch_size的样本个数打印一次, i*100iteration_loss = loss.item() / inputs.size(0)iteration_acc = 100 * torch.sum(preds == labels.data).item() / len(preds)print('Mode {}, Epoch [{}/{}], Step [{}/{}], Accuracy: {}, Loss: {:.4f}'.format(phase, epoch + 1,num_epochs, i + 1,total_step,iteration_acc,iteration_loss))epoch_loss = running_loss / train_datalengths[phase]epoch_acc = running_corrects.double() / train_datalengths[phase]if phase == 'train':train_loss.append(epoch_loss)train_acc.append(epoch_acc)else:validation_loss.append(epoch_loss)validation_acc.append(epoch_acc)print('*' * 10)print('Mode: [{}], Loss: {:.4f}, Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))print('*' * 10)# deep copy the modelif phase == 'val' and epoch_acc > best_acc:best_acc = epoch_accbest_model_wts = copy.deepcopy(model.state_dict())print()time_elapsed = time.time() - sinceprint('*' * 10)print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))print('Best val Acc: {:4f}'.format(best_acc))print('*' * 10)# load best model weightsfinal_model = copy.deepcopy(model) # 最后得到的modelmodel.load_state_dict(best_model_wts) # 在验证集上最好的modeldraw_loss_acc(train_list=train_loss, validation_list=validation_loss, mode='Loss') # 绘制Loss图像draw_loss_acc(train_list=train_acc, validation_list=validation_acc, mode='Accuracy') # 绘制准确率图像return (model, final_model)开始训练，并保存模型
if __name__ == "__main__":# 模型初始化model = ConvNet(num_classes=num_classes).to(device)# 打印模型结构#print(model)"""ConvNet((layer1): Sequential((0): Conv2d(1, 16, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(2): ReLU()(3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))(layer2): Sequential((0): Conv2d(16, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(2): ReLU()(3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))(fc): Linear(in_features=1568, out_features=10, bias=True))"""# -------------------# Loss and optimizer# ------------------criterion = nn.CrossEntropyLoss()optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)# -------------# 进行模型的训练# -------------(best_model, final_model) = train_model(model=model, criterion=criterion, optimizer=optimizer,dataloaders=train_dataloaders, train_datalengths=train_datalengths,num_epochs=num_epochs)"""Epoch [1/3]----------Mode train, Epoch [1/3], Step [100/570], Accuracy: 96.0, Loss: 0.0012Mode train, Epoch [1/3], Step [200/570], Accuracy: 96.0, Loss: 0.0011Mode train, Epoch [1/3], Step [300/570], Accuracy: 100.0, Loss: 0.0004Mode train, Epoch [1/3], Step [400/570], Accuracy: 100.0, Loss: 0.0003Mode train, Epoch [1/3], Step [500/570], Accuracy: 96.0, Loss: 0.0010**********Mode: [train], Loss: 0.1472, Acc: 0.9579********************Mode: [val], Loss: 0.0653, Acc: 0.9810**********Epoch [2/3]----------Mode train, Epoch [2/3], Step [100/570], Accuracy: 99.0, Loss: 0.0005Mode train, Epoch [2/3], Step [200/570], Accuracy: 98.0, Loss: 0.0003Mode train, Epoch [2/3], Step [300/570], Accuracy: 98.0, Loss: 0.0003Mode train, Epoch [2/3], Step [400/570], Accuracy: 97.0, Loss: 0.0005Mode train, Epoch [2/3], Step [500/570], Accuracy: 98.0, Loss: 0.0004**********Mode: [train], Loss: 0.0470, Acc: 0.9853********************Mode: [val], Loss: 0.0411, Acc: 0.9890**********Epoch [3/3]----------Mode train, Epoch [3/3], Step [100/570], Accuracy: 99.0, Loss: 0.0006Mode train, Epoch [3/3], Step [200/570], Accuracy: 99.0, Loss: 0.0009Mode train, Epoch [3/3], Step [300/570], Accuracy: 98.0, Loss: 0.0004Mode train, Epoch [3/3], Step [400/570], Accuracy: 99.0, Loss: 0.0003Mode train, Epoch [3/3], Step [500/570], Accuracy: 100.0, Loss: 0.0002**********Mode: [train], Loss: 0.0348, Acc: 0.9890********************Mode: [val], Loss: 0.0432, Acc: 0.9867********************Training complete in 0m 32sBest val Acc: 0.989000**********"""torch.save(model, 'CNN_MNIST.pkl')训练完的 loss 和 acc 曲线
载入模型进行简单的测试，以下面这张 7 为例子
model = torch.load('CNN_MNIST.pkl')# print(test_dataset.data[0].shape) # torch.Size([28, 28])# print(test_dataset.targets[0]) # tensor(7)#unload = transforms.ToPILImage()img = unload(test_dataset.data[0])img.save("test_data_0.jpg")# 带入模型进行预测inputdata = test_dataset.data[0].view(1,1,28,28).float()/255inputdata = inputdata.to(device)outputs = model(inputdata)print(outputs)"""tensor([[ -7.4075, -3.4618, -1.5322, 1.1591, -10.6026, -5.5818, -17.4020,14.6429, -4.5378, 1.0360]], device='cuda:0',grad_fn=<AddmmBackward0>)"""print(torch.max(outputs, 1))"""torch.return_types.max(values=tensor([14.6429], device='cuda:0', grad_fn=<MaxBackward0>),indices=tensor([7], device='cuda:0'))"""OK，预测结果正常，为 7
2 控制输出的概率, 看输入是什么步骤
初始随机一张图片我们希望让分类中某个数的概率最大最后看一下这个网络认为什么样的图像是这个数字 # hook类的写法class SaveFeatures():"""注册hook和移除hook"""def __init__(self, module):self.hook = module.register_forward_hook(self.hook_fn)def hook_fn(self, module, input, output):# self.features = output.clone().detach().requires_grad_(True)self.features = output.clone()def close(self):self.hook.remove()下面开始操作
# hook住模型layer = 2activations = SaveFeatures(list(model.children())[layer])# 超参数lr = 0.005 # 学习率opt_steps = 100 # 迭代次数upscaling_factor = 10 # 放大的倍数(为了最后图像的保存)# 保存迭代后的数字true_nums = []random_init, num_init = 1, 0 # the type of initial# 带入网络进行迭代for true_num in range(0, 10):# 初始化随机图片(数据定义和优化器一定要在一起)# 定义数据sz = 28if random_init:img = np.uint8(np.random.uniform(0, 255, (1,sz, sz))) / 255img = torch.from_numpy(img[None]).float().to(device)img_var = Variable(img, requires_grad=True)if num_init:# 将7变成0,1,2,3,4,5,6,7,8,9img = test_dataset.data[0].view(1, 1, 28, 28).float() / 255img = img.to(device)img_var = Variable(img, requires_grad=True)# 定义优化器optimizer = torch.optim.Adam([img_var], lr=lr, weight_decay=1e-6)for n in range(opt_steps): # optimize pixel values for opt_steps timesoptimizer.zero_grad()model(img_var) # 正向传播if random_init:loss = -activations.features[0, true_num] # 这里的loss确保某个值的输出大if num_init:loss = -activations.features[0, true_num] + F.mse_loss(img_var, img) # 这里的loss确保某个值的输出大, 并且与原图不会相差很多loss.backward()optimizer.step()# 打印最后的img的样子print(activations.features[0, true_num]) # tensor(23.8736, device='cuda:0', grad_fn=<SelectBackward0>)print(activations.features[0])"""tensor([ 23.8736, -32.0724, -5.7329, -18.6501, -16.1558, -18.9483, -0.3033,-29.1561, 14.9260, -13.5412], device='cuda:0',grad_fn=<SelectBackward0>)"""print('========')img = img_var.cpu().clone()img = img.squeeze(0)# 图像的裁剪(确保像素值的范围)img[img > 1] = 1img[img < 0] = 0true_nums.append(img)unloader = transforms.ToPILImage()img = unloader(img)img = cv2.cvtColor(np.asarray(img), cv2.COLOR_RGB2BGR)sz = int(upscaling_factor * sz) # calculate new image sizeimg = cv2.resize(img, (sz, sz), interpolation=cv2.INTER_CUBIC) # scale image upif random_init:cv2.imwrite('random_regonize{}.jpg'.format(true_num), img)if num_init:cv2.imwrite('real7_regonize{}.jpg'.format(true_num), img)"""========tensor(22.3530, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-22.4025, 22.3530, 6.0451, -9.4460, -1.0577, -17.7650, -11.3686,-11.8474, -5.5310, -16.9936], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(50.2202, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-16.9364, -16.5120, 50.2202, -9.5287, -25.2837, -32.3480, -22.8569,-20.1231, 1.1174, -31.7244], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(48.8004, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-33.3715, -30.5732, -6.0252, 48.8004, -34.9467, -17.8136, -35.1371,-17.4484, 2.8954, -11.9694], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(31.5068, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-24.5204, -13.6857, -5.1833, -22.7889, 31.5068, -20.2855, -16.7245,-19.1719, 2.4699, -16.2246], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(37.4866, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-20.2235, -26.1013, -25.9511, -2.7806, -19.5546, 37.4866, -10.8689,-30.3888, 0.1591, -8.5250], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(35.9310, device='cuda:0', grad_fn=<SelectBackward0>)tensor([ -6.1996, -31.2246, -8.3396, -21.6307, -16.9098, -9.5194, 35.9310,-33.0918, 10.2462, -28.6393], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(23.6772, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-21.5441, -10.3366, -4.9905, -0.9289, -6.9219, -23.5643, -23.9894,23.6772, -0.7960, -16.9556], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(57.8378, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-13.5191, -53.9004, -9.2996, -10.3597, -27.3806, -27.5858, -15.3235,-46.7014, 57.8378, -17.2299], device='cuda:0',grad_fn=<SelectBackward0>)========tensor(37.0334, device='cuda:0', grad_fn=<SelectBackward0>)tensor([-26.2983, -37.7131, -16.6210, -1.8686, -11.5330, -11.7843, -25.7539,-27.0036, 6.3785, 37.0334], device='cuda:0',grad_fn=<SelectBackward0>)========"""# 移除hookactivations.close()for i in range(0,10):_ , pre = torch.max(model(true_nums[i][None].to(device)),1)print("i:{},Pre:{}".format(i,pre))"""i:0,Pre:tensor([0], device='cuda:0')i:1,Pre:tensor([1], device='cuda:0')i:2,Pre:tensor([2], device='cuda:0')i:3,Pre:tensor([3], device='cuda:0')i:4,Pre:tensor([4], device='cuda:0')i:5,Pre:tensor([5], device='cuda:0')i:6,Pre:tensor([6], device='cuda:0')i:7,Pre:tensor([7], device='cuda:0')i:8,Pre:tensor([8], device='cuda:0')i:9,Pre:tensor([9], device='cuda:0')"""可以看到相应最高的图片的样子，依稀可以看见 0，1，2，3，4，5，6，7，8，9 的轮廓
3 让正确的图片分类错误步骤如下
使用特定数字的图片，如数字7(初始化方式与前面的不一样，是特定而不是随机)使用上面训练好的网络，固定网络参数；最后一层因为是10个输出(相当于是概率), 我们 loss 设置为某个的负的概率值（利于梯度下降）梯度下降, 使负的概率下降，即相对应的概率值上升，我们最终修改的是初始的图片这样使得网络认为这张图片识别的数字的概率增加简单描述，输入图片数字7，固定网络参数，让其他数字的概率不断变大，改变输入图片，最终达到迷惑网络的目的
只要将第二节中的代码 random_init, num_init = 1, 0 # the type of initial 改为 random_init, num_init = 0, 1 # the type of initial 即可
成功用 7 迷惑了所有类别
注意损失函数中还额外引入了 F.mse_loss(img_var, img) ，确保某个数字类别的概率输出增大, 并且与原图数字 7 像素上不会相差很多
