efficient_weight/utils/attacks.py

import os 
import argparse
import torch
import torch.nn as nn 
import torch.nn.functional as F
import torchvision
from torch.autograd import Variable
import torch.optim as optim
from torchvision import datasets, transforms
import numpy as np 

def PGD_Revise(model, img, gt, epsilon, step_size, num_steps, device):
        model.eval()
        x_adv = img.detach() + torch.from_numpy(np.random.uniform(-epsilon, epsilon, img.shape)).float().to(device)
        x_adv = torch.clamp(x_adv, 0.0, 1.0)
        for k in range(num_steps):
            x_adv.requires_grad_()
            output = model(x_adv)
            model.zero_grad()
            with torch.enable_grad():
                loss_adv = nn.CrossEntropyLoss()(output, gt)
                loss_adv.backward()
            eta = step_size * x_adv.grad.sign()
            # Update adversarial img
            x_adv = x_adv.detach() + eta
            x_adv = torch.min(torch.max(x_adv, img - epsilon), img + epsilon)
            x_adv = torch.clamp(x_adv, 0.0, 1.0)
            if k == (num_steps//2) :
                x_adv_mid = x_adv.detach().clone()
        x_adv = Variable(x_adv, requires_grad=False)
        x_adv_mid = Variable(x_adv_mid, requires_grad = False)
        return x_adv, x_adv_mid 
    
    
def PGD_Revise2(model, img, gt, epsilon, step_size, num_steps, device):
        model.eval()
        x_adv = img.detach() + torch.from_numpy(np.random.uniform(-epsilon, epsilon, img.shape)).float().to(device)
        x_adv = torch.clamp(x_adv, 0.0, 1.0)
        for k in range(num_steps):
            x_adv.requires_grad_()
            output = model(x_adv)
            model.zero_grad()
            with torch.enable_grad():
                loss_adv = nn.CrossEntropyLoss()(output, gt)
                loss_adv.backward()
            eta = step_size * x_adv.grad.sign()
            # Update adversarial img
            x_adv = x_adv.detach() + eta
            x_adv = torch.min(torch.max(x_adv, img - epsilon), img + epsilon)
            x_adv = torch.clamp(x_adv, 0.0, 1.0)
            if k == (num_steps//2 -1) :
                x_adv_now = x_adv.detach().clone()
        x_adv = Variable(x_adv, requires_grad=False)
        x_adv_now = Variable(x_adv_now, requires_grad = False)
        return x_adv_now, x_adv