【对抗攻击系列_PGD_2018_ICLR】Towards Deep Learning Models Resistant to Adversarial Attacks

class PGD(nn.Module):
    def __init__(self,model):
        super().__init__()
        self.model=model#必须是pytorch的model
        self.device=torch.device("cuda" if (torch.cuda.is_available()) else "cpu")
    def generate(self,x,**params):
        self.parse_params(**params)
        labels=self.y

        adv_x=self.attack(x,labels)
        return adv_x
    def parse_params(self,eps=0.3,iter_eps=0.01,nb_iter=40,clip_min=0.0,clip_max=1.0,C=0.0,
                     y=None,ord=np.inf,rand_init=True,flag_target=False):
        self.eps=eps
        self.iter_eps=iter_eps
        self.nb_iter=nb_iter
        self.clip_min=clip_min
        self.clip_max=clip_max
        self.y=y
        self.ord=ord
        self.rand_init=rand_init
        self.model.to(self.device)
        self.flag_target=flag_target
        self.C=C


    def sigle_step_attack(self,x,pertubation,labels):
        adv_x=x+pertubation
        # get the gradient of x
        adv_x=Variable(adv_x)
        adv_x.requires_grad = True
        loss_func=nn.CrossEntropyLoss()
        preds=self.model(adv_x)
        if self.flag_target:
            loss =-loss_func(preds,labels)
        else:
            loss=loss_func(preds,labels)
            # label_mask=torch_one_hot(labels)
            #
            # correct_logit=torch.mean(torch.sum(label_mask * preds,dim=1))
            # wrong_logit = torch.mean(torch.max((1 - label_mask) * preds, dim=1)[0])
            # loss=-F.relu(correct_logit-wrong_logit+self.C)

        self.model.zero_grad()
        loss.backward()
        grad=adv_x.grad.data
        #get the pertubation of an iter_eps
        pertubation=self.iter_eps*np.sign(grad)
        adv_x=adv_x.cpu().detach().numpy()+pertubation.cpu().numpy()
        x=x.cpu().detach().numpy()

        pertubation=np.clip(adv_x,self.clip_min,self.clip_max)-x
        pertubation=clip_pertubation(pertubation,self.ord,self.eps)


        return pertubation
    def attack(self,x,labels):
        labels = labels.to(self.device)
        print(self.rand_init)
        if self.rand_init:
            x_tmp=x+torch.Tensor(np.random.uniform(-self.eps, self.eps, x.shape)).type_as(x).cuda()
        else:
            x_tmp=x
        pertubation=torch.zeros(x.shape).type_as(x).to(self.device)
        for i in range(self.nb_iter):
            pertubation=self.sigle_step_attack(x_tmp,pertubation=pertubation,labels=labels)
            pertubation=torch.Tensor(pertubation).type_as(x).to(self.device)
        adv_x=x+pertubation
        adv_x=adv_x.cpu().detach().numpy()

        adv_x=np.clip(adv_x,self.clip_min,self.clip_max)

        return adv_x