青岛专业网站制作设计,福建电信网站备案,网站备案审核制度,wordpress附件页面有现成的库cleverhans可以调用
https://github.com/cleverhans-lab/cleverhans
这个库是用tensorflow实现的#xff0c;但也实现了pytorch版本
因为我调用的是pytorch版本#xff0c;我接下来大概讲一下如何使用
他们给出了一个调用pytorch版本的例子#xff1a;
https…有现成的库cleverhans可以调用
https://github.com/cleverhans-lab/cleverhans
这个库是用tensorflow实现的但也实现了pytorch版本
因为我调用的是pytorch版本我接下来大概讲一下如何使用
他们给出了一个调用pytorch版本的例子
https://github.com/cleverhans-lab/cleverhans/blob/master/tutorials/torch/cifar10_tutorial.py
代码如下
from absl import app, flags
from easydict import EasyDict
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvisionfrom cleverhans.torch.attacks.fast_gradient_method import fast_gradient_method
from cleverhans.torch.attacks.projected_gradient_descent import projected_gradient_descentFLAGS flags.FLAGSclass CNN(torch.nn.Module):Basic CNN architecture.def __init__(self, in_channels1):super(CNN, self).__init__()self.conv1 nn.Conv2d(in_channels, 64, 8, 1)self.conv2 nn.Conv2d(64, 128, 6, 2)self.conv3 nn.Conv2d(128, 128, 5, 2)self.fc nn.Linear(128*3*3, 10)def forward(self, x):x F.relu(self.conv1(x))x F.relu(self.conv2(x))x F.relu(self.conv3(x))x x.view(-1, 128*3*3)x self.fc(x)return xdef ld_cifar10():Load training and test data.train_transforms torchvision.transforms.Compose([torchvision.transforms.ToTensor()])test_transforms torchvision.transforms.Compose([torchvision.transforms.ToTensor()])train_dataset torchvision.datasets.CIFAR10(root/tmp/data, trainTrue, transformtrain_transforms, downloadTrue)test_dataset torchvision.datasets.CIFAR10(root/tmp/data, trainFalse, transformtest_transforms, downloadTrue)train_loader torch.utils.data.DataLoader(train_dataset, batch_size128, shuffleTrue, num_workers2)test_loader torch.utils.data.DataLoader(test_dataset, batch_size128, shuffleFalse, num_workers2)return EasyDict(traintrain_loader, testtest_loader)def main(_):# Load training and test datadata ld_cifar10()# Instantiate model, loss, and optimizer for trainingnet CNN(in_channels3)device cuda if torch.cuda.is_available() else cpuif device cuda:net net.cuda()loss_fn torch.nn.CrossEntropyLoss(reductionmean)optimizer torch.optim.Adam(net.parameters(), lr1e-3)# Train vanilla modelnet.train()for epoch in range(1, FLAGS.nb_epochs 1):train_loss 0.for x, y in data.train:x, y x.to(device), y.to(device)if FLAGS.adv_train:# Replace clean example with adversarial example for adversarial trainingx projected_gradient_descent(net, x, FLAGS.eps, 0.01, 40, np.inf)optimizer.zero_grad()loss loss_fn(net(x), y)loss.backward()optimizer.step()train_loss loss.item()print(epoch: {}/{}, train loss: {:.3f}.format(epoch, FLAGS.nb_epochs, train_loss))# Evaluate on clean and adversarial datanet.eval()report EasyDict(nb_test0, correct0, correct_fgm0, correct_pgd0)for x, y in data.test:x, y x.to(device), y.to(device)x_fgm fast_gradient_method(net, x, FLAGS.eps, np.inf)x_pgd projected_gradient_descent(net, x, FLAGS.eps, 0.01, 40, np.inf)_, y_pred net(x).max(1) # model prediction on clean examples_, y_pred_fgm net(x_fgm).max(1) # model prediction on FGM adversarial examples_, y_pred_pgd net(x_pgd).max(1) # model prediction on PGD adversarial examplesreport.nb_test y.size(0)report.correct y_pred.eq(y).sum().item()report.correct_fgm y_pred_fgm.eq(y).sum().item()report.correct_pgd y_pred_pgd.eq(y).sum().item()print(test acc on clean examples (%): {:.3f}.format(report.correct / report.nb_test * 100.))print(test acc on FGM adversarial examples (%): {:.3f}.format(report.correct_fgm / report.nb_test * 100.))print(test acc on PGD adversarial examples (%): {:.3f}.format(report.correct_pgd / report.nb_test * 100.))if __name__ __main__:flags.DEFINE_integer(nb_epochs, 8, Number of epochs.)flags.DEFINE_float(eps, 0.3, Total epsilon for FGM and PGD attacks.)flags.DEFINE_bool(adv_train, False, Use adversarial training (on PGD adversarial examples).)app.run(main)
代码不细讲关键点在于下面这句通过这句利用FGSM生成对抗样本
x_fgm fast_gradient_method(net, x, FLAGS.eps, np.inf)
有一点需要注意上述代码中需要引入了
from cleverhans.torch.attacks.fast_gradient_method import fast_gradient_method
github的readme中中说首先需要安装
pip install cleverhans
我确实可以安装这个库导入也没有问题但 执行到下面这句就是不通过
from cleverhans.torch.attacks.fast_gradient_method import fast_gradient_method
说cleverhans.torch不存在
于是没办法我只能把github上实现的代码复制下来自己建一个一模一样的fast_gradient_method方法
