50次pgd攻击下生成mnist数据集的对抗样本，并将对抗样本转化为npy文件

我们先定义PGD攻击的函数：

import torch
import numpy as np


def pgd_attack(model, X, y, epsilon, alpha, num_iter):
    """PGD攻击（Projected Gradient Descent）"""
    delta = torch.zeros_like(X, requires_grad=True)
    for t in range(num_iter):
        loss = torch.nn.CrossEntropyLoss()(model(X + delta), y)
        loss.backward()
        delta.data = (delta + alpha * delta.grad.detach().sign()).clamp(-epsilon, epsilon)
        delta.grad.zero_()
    return (X + delta).detach()

接下来，我们加载MNIST数据集和预训练模型：

import torchvision
from torchvision import transforms


transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])
testset = torchvision.datasets.MNIST(root='./data', train=False, download=True, transform=transform)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = torch.load('mnist_cnn.pth').to(device)

然后我们可以用PGD攻击来生成对抗样本：

epsilon = 0.3
alpha = 0.05
num_iter = 50

adv_samples = []
for i, (X, y) in enumerate(testset):
    X, y = X.unsqueeze(0).to(device), y.unsqueeze(0).to(device)
    adv = pgd_attack(model, X, y, epsilon, alpha, num_iter)
    adv_samples.append(adv.cpu().numpy().squeeze())
    if i % 1000 == 0:
        print(f"Generated {i} adv samples")

adv_samples = np.asarray(adv_samples)

最后，我们将对抗样本保存为`npy`文件：

np.save('adv_samples.npy', adv_samples)

完整代码如下：