Adversarial-Machine-Learning-Clinic/Filter_Analysis/fgsm.py

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
import cv2
from mnist import Net
from pykuwahara import kuwahara

epsilons = np.arange(0.05,0.35,0.05)
pretrained_model = "mnist_cnn_unfiltered.pt"
use_cuda=False

torch.manual_seed(69)


test_loader = torch.utils.data.DataLoader(
        datasets.MNIST('data/', train=False, download=True, transform=transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.1307,), (0.3081,)),
            ])),
        batch_size=1, shuffle=True)

print("CUDA Available: ", torch.cuda.is_available())
device = torch.device("cuda" if use_cuda and torch.cuda.is_available() else "cpu")

model = Net().to(device)
print(type(model))

model.load_state_dict(torch.load(pretrained_model, map_location=device))

model.eval()

def fgsm_attack(image, epsilon, data_grad):
    # Collect the element-wise sign of the data gradient
    sign_data_grad = data_grad.sign()
    
    # Create the perturbed image by adjusting each pixel of the input image
    perturbed_image = image + epsilon*sign_data_grad

    # Adding clipping to maintain [0, 1] range
    perturbed_image = torch.clamp(perturbed_image, 0, 1)
    
    return perturbed_image

def denorm(batch, mean=[0.1307], std=[0.3081]):
    """
    Convert a batch of tensors to their original scale.

    Args:
        batch (torch.Tensor): Batch of normalized tensors.
        mean (torch.Tensor or list): Man used for normalization.
        std (torch.Tensor or list): Standard deviation used for normalization.

    Returns:
        torch.Tensor: batch of tensors without normalization applied to them.
    """

    if isinstance(mean, list):
        mean = torch.tensor(mean).to(device)
    if isinstance(std, list):
        std = torch.tensor(std).to(device)

    return batch * std.view(1, -1, 1, 1) + mean.view(1, -1, 1, 1)

def test(model, device, test_loader, epsilon):
    # Original dataset correct classifications
    orig_correct = 0
    # Attacked dataset correct classifications
    attacked_correct = 0
    # Filtered attacked dataset correct classifications
    filtered_correct = 0

    adv_examples = []

    for data, target in test_loader:
        data, target = data.to(device), target.to(device)
        data.requires_grad = True

        output_orig = model(data)
        orig_pred = output_orig.max(1, keepdim=True)[1]

        # Calculate the loss
        loss = F.nll_loss(output_orig, target)

        # Zero all existing gradients
        model.zero_grad()

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

        # Collect ''datagrad''
        data_grad = data.grad.data

        # Restore the data to its original scale
        data_denorm = denorm(data)

        # Apply the FGSM attack
        perturbed_data = fgsm_attack(data_denorm, epsilon, data_grad)

        # Reapply normalization
        perturbed_data_normalized = transforms.Normalize((0.1307,), (0.3081,))(perturbed_data)

        # Filter the attacked image
        perturbed_data_filtered = filtered(perturbed_data_normalized, len(perturbed_data_normalized))

        # evaluate the model on the attacked and filtered images
        output_attacked = model(perturbed_data_normalized)
        output_filtered = model(perturbed_data_filtered)

        attacked_pred = output_attacked.max(1, keepdim=True)[1]
        filtered_pred = output_filtered.max(1, keepdim=True)[1]

        if orig_pred.item() == target.item():
            orig_correct += 1

        if attacked_pred.item() == target.item():
            attacked_correct += 1

            if epsilon == 0 and len(adv_examples) < 5:
                adv_ex = perturbed_data.squeeze().detach().cpu().numpy()
                adv_examples.append( (orig_pred.item(), attacked_pred.item(), adv_ex) )

        if filtered_pred.item() == target.item():
            filtered_correct += 1

            if epsilon == 0 and len(adv_examples) < 5:
                adv_ex = perturbed_data.squeeze().detach().cpu().numpy()
                adv_examples.append( (orig_pred.item(), filtered_pred.item(), adv_ex) )

    orig_acc = orig_correct/float(len(test_loader))
    attacked_acc = attacked_correct/float(len(test_loader))
    filtered_acc = filtered_correct/float(len(test_loader))
    print(f"Epsilon: {epsilon}")
    print(f"Original Accuracy = {orig_correct} / {len(test_loader)} = {orig_acc}")
    print(f"Attacked Accuracy = {attacked_correct} / {len(test_loader)} = {attacked_acc}")
    print(f"Filtered Accuracy = {filtered_correct} / {len(test_loader)} = {filtered_acc}")
    print(f"Filtered:Attacked = {filtered_acc} / {attacked_acc} = {filtered_acc/attacked_acc}")

    return attacked_acc, filtered_acc, adv_examples


def filtered(data, batch_size=64):
    # Turn the tensor into an image
    images = None
    try:
        images = data.numpy().transpose(0,2,3,1)
    except RuntimeError:
        images = data.detach().numpy().transpose(0,2,3,1)

    # Apply the Kuwahara filter
    filtered_images = np.ndarray((batch_size,28,28,1))

    for i in range(batch_size):
        filtered_images[i] = kuwahara(images[i], method='gaussian', radius=5, image_2d=images[i])

    # Modify the data with the filtered image
    filtered_images = filtered_images.transpose(0,3,1,2)
    return torch.tensor(filtered_images).float()

attacked_accuracies = []
filtered_accuracies = []
ratios = []
examples = []

print(f"Model: {pretrained_model}")

for eps in epsilons:
    aacc, facc, ex = test(model, device, test_loader, eps)
    attacked_accuracies.append(aacc)
    filtered_accuracies.append(facc)
    ratios.append(facc/aacc)
    examples.append(ex)

# Plot the results
plt.subplot(121)
plt.plot(epsilons, attacked_accuracies, label="Attacked Accuracy")
plt.plot(epsilons, filtered_accuracies, label="Filtered Accuracy")
plt.legend()

plt.subplot(122)
plt.plot(epsilons, ratios, label="Filtered:Attacked")
plt.legend()
plt.show()