SMART/adversarial-attacks-pytorch/torchattacks/attack.py

import time
from collections import OrderedDict

import torch
from torch.utils.data import DataLoader, TensorDataset


def wrapper_method(func):
    def wrapper_func(self, *args, **kwargs):
        result = func(self, *args, **kwargs)
        for atk in self.__dict__.get("_attacks").values():
            eval("atk." + func.__name__ + "(*args, **kwargs)")
        return result

    return wrapper_func


class Attack(object):
    r"""
    Base class for all attacks.

    .. note::
        It automatically set device to the device where given model is.
        It basically changes training mode to eval during attack process.
        To change this, please see `set_model_training_mode`.
    """

    def __init__(self, name, model):
        r"""
        Initializes internal attack state.

        Arguments:
            name (str): name of attack.
            model (torch.nn.Module): model to attack.
        """

        self.attack = name
        self._attacks = OrderedDict()

        self.set_model(model)
        try:
            self.device = next(model.parameters()).device
        except Exception:
            self.device = None
            print("Failed to set device automatically, please try set_device() manual.")

        # Controls attack mode.
        self.attack_mode = "default"
        self.supported_mode = ["default"]
        self.targeted = False
        self._target_map_function = None

        # Controls when normalization is used.
        self.normalization_used = None
        self._normalization_applied = None
        if self.model.__class__.__name__ == "RobModel":
            self._set_rmodel_normalization_used(model)

        # Controls model mode during attack.
        self._model_training = False
        self._batchnorm_training = False
        self._dropout_training = False

    def forward(self, inputs, labels=None, *args, **kwargs):
        r"""
        It defines the computation performed at every call.
        Should be overridden by all subclasses.
        """
        raise NotImplementedError

    @wrapper_method
    def set_model(self, model):
        self.model = model
        self.model_name = model.__class__.__name__

    def get_logits(self, inputs, labels=None, *args, **kwargs):
        if self._normalization_applied is False:
            inputs = self.normalize(inputs)
        logits = self.model(inputs)
        return logits

    @wrapper_method
    def _set_normalization_applied(self, flag):
        self._normalization_applied = flag

    @wrapper_method
    def set_device(self, device):
        self.device = device

    @wrapper_method
    def _set_rmodel_normalization_used(self, model):
        r"""
        Set attack normalization for MAIR [https://github.com/Harry24k/MAIR].

        """
        mean = getattr(model, "mean", None)
        std = getattr(model, "std", None)
        if (mean is not None) and (std is not None):
            if isinstance(mean, torch.Tensor):
                mean = mean.cpu().numpy()
            if isinstance(std, torch.Tensor):
                std = std.cpu().numpy()
            if (mean != 0).all() or (std != 1).all():
                self.set_normalization_used(mean, std)

    @wrapper_method
    def set_normalization_used(self, mean, std):
        self.normalization_used = {}
        n_channels = len(mean)
        mean = torch.tensor(mean).reshape(1, n_channels, 1, 1)
        std = torch.tensor(std).reshape(1, n_channels, 1, 1)
        self.normalization_used["mean"] = mean
        self.normalization_used["std"] = std
        self._set_normalization_applied(True)

    def normalize(self, inputs):
        mean = self.normalization_used["mean"].to(inputs.device)
        std = self.normalization_used["std"].to(inputs.device)
        return (inputs - mean) / std

    def inverse_normalize(self, inputs):
        mean = self.normalization_used["mean"].to(inputs.device)
        std = self.normalization_used["std"].to(inputs.device)
        return inputs * std + mean

    def get_mode(self):
        r"""
        Get attack mode.

        """
        return self.attack_mode

    @wrapper_method
    def set_mode_default(self):
        r"""
        Set attack mode as default mode.

        """
        self.attack_mode = "default"
        self.targeted = False
        print("Attack mode is changed to 'default.'")

    @wrapper_method
    def _set_mode_targeted(self, mode, quiet):
        if "targeted" not in self.supported_mode:
            raise ValueError("Targeted mode is not supported.")
        self.targeted = True
        self.attack_mode = mode
        if not quiet:
            print("Attack mode is changed to '%s'." % mode)

    @wrapper_method
    def set_mode_targeted_by_function(self, target_map_function, quiet=False):
        r"""
        Set attack mode as targeted.

        Arguments:
            target_map_function (function): Label mapping function.
                e.g. lambda inputs, labels:(labels+1)%10.
                None for using input labels as targeted labels. (Default)
            quiet (bool): Display information message or not. (Default: False)

        """
        self._set_mode_targeted("targeted(custom)", quiet)
        self._target_map_function = target_map_function

    @wrapper_method
    def set_mode_targeted_random(self, quiet=False):
        r"""
        Set attack mode as targeted with random labels.

        Arguments:
            quiet (bool): Display information message or not. (Default: False)

        """
        self._set_mode_targeted("targeted(random)", quiet)
        self._target_map_function = self.get_random_target_label

    @wrapper_method
    def set_mode_targeted_least_likely(self, kth_min=1, quiet=False):
        r"""
        Set attack mode as targeted with least likely labels.

        Arguments:
            kth_min (str): label with the k-th smallest probability used as target labels. (Default: 1)
            num_classses (str): number of classes. (Default: False)

        """
        self._set_mode_targeted("targeted(least-likely)", quiet)
        assert kth_min > 0
        self._kth_min = kth_min
        self._target_map_function = self.get_least_likely_label

    @wrapper_method
    def set_mode_targeted_by_label(self, quiet=False):
        r"""
        Set attack mode as targeted.

        Arguments:
            quiet (bool): Display information message or not. (Default: False)

        .. note::
            Use user-supplied labels as target labels.
        """
        self._set_mode_targeted("targeted(label)", quiet)
        self._target_map_function = "function is a string"

    @wrapper_method
    def set_model_training_mode(
        self, model_training=False, batchnorm_training=False, dropout_training=False
    ):
        r"""
        Set training mode during attack process.

        Arguments:
            model_training (bool): True for using training mode for the entire model during attack process.
            batchnorm_training (bool): True for using training mode for batchnorms during attack process.
            dropout_training (bool): True for using training mode for dropouts during attack process.

        .. note::
            For RNN-based models, we cannot calculate gradients with eval mode.
            Thus, it should be changed to the training mode during the attack.
        """
        self._model_training = model_training
        self._batchnorm_training = batchnorm_training
        self._dropout_training = dropout_training

    @wrapper_method
    def _change_model_mode(self, given_training):
        if self._model_training:
            self.model.train()
            for _, m in self.model.named_modules():
                if not self._batchnorm_training:
                    if "BatchNorm" in m.__class__.__name__:
                        m = m.eval()
                if not self._dropout_training:
                    if "Dropout" in m.__class__.__name__:
                        m = m.eval()
        else:
            self.model.eval()

    @wrapper_method
    def _recover_model_mode(self, given_training):
        if given_training:
            self.model.train()

    def save(
        self,
        data_loader,
        save_path=None,
        verbose=True,
        return_verbose=False,
        save_predictions=False,
        save_clean_inputs=False,
        save_type="float",
    ):
        r"""
        Save adversarial inputs as torch.tensor from given torch.utils.data.DataLoader.

        Arguments:
            save_path (str): save_path.
            data_loader (torch.utils.data.DataLoader): data loader.
            verbose (bool): True for displaying detailed information. (Default: True)
            return_verbose (bool): True for returning detailed information. (Default: False)
            save_predictions (bool): True for saving predicted labels (Default: False)
            save_clean_inputs (bool): True for saving clean inputs (Default: False)

        """
        if save_path is not None:
            adv_input_list = []
            label_list = []
            if save_predictions:
                pred_list = []
            if save_clean_inputs:
                input_list = []

        correct = 0
        total = 0
        l2_distance = []

        total_batch = len(data_loader)
        given_training = self.model.training

        for step, (inputs, labels) in enumerate(data_loader):
            start = time.time()
            adv_inputs = self.__call__(inputs, labels)
            batch_size = len(inputs)

            if verbose or return_verbose:
                with torch.no_grad():
                    outputs = self.get_output_with_eval_nograd(adv_inputs)

                    # Calculate robust accuracy
                    _, pred = torch.max(outputs.data, 1)
                    total += labels.size(0)
                    right_idx = pred == labels.to(self.device)
                    correct += right_idx.sum()
                    rob_acc = 100 * float(correct) / total

                    # Calculate l2 distance
                    delta = (adv_inputs - inputs.to(self.device)).view(
                        batch_size, -1
                    )  # nopep8
                    l2_distance.append(
                        torch.norm(delta[~right_idx], p=2, dim=1)
                    )  # nopep8
                    l2 = torch.cat(l2_distance).mean().item()

                    # Calculate time computation
                    progress = (step + 1) / total_batch * 100
                    end = time.time()
                    elapsed_time = end - start

                    if verbose:
                        self._save_print(
                            progress, rob_acc, l2, elapsed_time, end="\r"
                        )  # nopep8

            if save_path is not None:
                adv_input_list.append(adv_inputs.detach().cpu())
                label_list.append(labels.detach().cpu())

                adv_input_list_cat = torch.cat(adv_input_list, 0)
                label_list_cat = torch.cat(label_list, 0)

                save_dict = {
                    "adv_inputs": adv_input_list_cat,
                    "labels": label_list_cat,
                }  # nopep8

                if save_predictions:
                    pred_list.append(pred.detach().cpu())
                    pred_list_cat = torch.cat(pred_list, 0)
                    save_dict["preds"] = pred_list_cat

                if save_clean_inputs:
                    input_list.append(inputs.detach().cpu())
                    input_list_cat = torch.cat(input_list, 0)
                    save_dict["clean_inputs"] = input_list_cat

                if self.normalization_used is not None:
                    save_dict["adv_inputs"] = self.inverse_normalize(
                        save_dict["adv_inputs"]
                    )  # nopep8
                    if save_clean_inputs:
                        save_dict["clean_inputs"] = self.inverse_normalize(
                            save_dict["clean_inputs"]
                        )  # nopep8

                if save_type == "int":
                    save_dict["adv_inputs"] = self.to_type(
                        save_dict["adv_inputs"], "int"
                    )  # nopep8
                    if save_clean_inputs:
                        save_dict["clean_inputs"] = self.to_type(
                            save_dict["clean_inputs"], "int"
                        )  # nopep8

                save_dict["save_type"] = save_type
                torch.save(save_dict, save_path)

        # To avoid erasing the printed information.
        if verbose:
            self._save_print(progress, rob_acc, l2, elapsed_time, end="\n")

        if given_training:
            self.model.train()

        if return_verbose:
            return rob_acc, l2, elapsed_time

    @staticmethod
    def to_type(inputs, type):
        r"""
        Return inputs as int if float is given.
        """
        if type == "int":
            if isinstance(inputs, torch.FloatTensor) or isinstance(
                inputs, torch.cuda.FloatTensor
            ):
                return (inputs * 255).type(torch.uint8)
        elif type == "float":
            if isinstance(inputs, torch.ByteTensor) or isinstance(
                inputs, torch.cuda.ByteTensor
            ):
                return inputs.float() / 255
        else:
            raise ValueError(type + " is not a valid type. [Options: float, int]")
        return inputs

    @staticmethod
    def _save_print(progress, rob_acc, l2, elapsed_time, end):
        print(
            "- Save progress: %2.2f %% / Robust accuracy: %2.2f %% / L2: %1.5f (%2.3f it/s) \t"
            % (progress, rob_acc, l2, elapsed_time),
            end=end,
        )

    @staticmethod
    def load(
        load_path,
        batch_size=128,
        shuffle=False,
        normalize=None,
        load_predictions=False,
        load_clean_inputs=False,
    ):
        save_dict = torch.load(load_path)
        keys = ["adv_inputs", "labels"]

        if load_predictions:
            keys.append("preds")
        if load_clean_inputs:
            keys.append("clean_inputs")

        if save_dict["save_type"] == "int":
            save_dict["adv_inputs"] = save_dict["adv_inputs"].float() / 255
            if load_clean_inputs:
                save_dict["clean_inputs"] = (
                    save_dict["clean_inputs"].float() / 255
                )  # nopep8

        if normalize is not None:
            n_channels = len(normalize["mean"])
            mean = torch.tensor(normalize["mean"]).reshape(1, n_channels, 1, 1)
            std = torch.tensor(normalize["std"]).reshape(1, n_channels, 1, 1)
            save_dict["adv_inputs"] = (save_dict["adv_inputs"] - mean) / std
            if load_clean_inputs:
                save_dict["clean_inputs"] = (
                    save_dict["clean_inputs"] - mean
                ) / std  # nopep8

        adv_data = TensorDataset(*[save_dict[key] for key in keys])
        adv_loader = DataLoader(adv_data, batch_size=batch_size, shuffle=shuffle)
        print(
            "Data is loaded in the following order: [%s]" % (", ".join(keys))
        )  # nopep8
        return adv_loader

    @torch.no_grad()
    def get_output_with_eval_nograd(self, inputs):
        given_training = self.model.training
        if given_training:
            self.model.eval()
        outputs = self.get_logits(inputs)
        if given_training:
            self.model.train()
        return outputs

    def get_target_label(self, inputs, labels=None):
        r"""
        Function for changing the attack mode.
        Return input labels.
        """
        if self._target_map_function is None:
            raise ValueError(
                "target_map_function is not initialized by set_mode_targeted."
            )
        if self.attack_mode == "targeted(label)":
            target_labels = labels
        else:
            target_labels = self._target_map_function(inputs, labels)
        return target_labels

    @torch.no_grad()
    def get_least_likely_label(self, inputs, labels=None):
        outputs = self.get_output_with_eval_nograd(inputs)
        if labels is None:
            _, labels = torch.max(outputs, dim=1)
        n_classses = outputs.shape[-1]

        target_labels = torch.zeros_like(labels)
        for counter in range(labels.shape[0]):
            l = list(range(n_classses))
            l.remove(labels[counter])
            _, t = torch.kthvalue(outputs[counter][l], self._kth_min)
            target_labels[counter] = l[t]

        return target_labels.long().to(self.device)

    @torch.no_grad()
    def get_random_target_label(self, inputs, labels=None):
        outputs = self.get_output_with_eval_nograd(inputs)
        if labels is None:
            _, labels = torch.max(outputs, dim=1)
        n_classses = outputs.shape[-1]

        target_labels = torch.zeros_like(labels)
        for counter in range(labels.shape[0]):
            l = list(range(n_classses))
            l.remove(labels[counter])
            t = (len(l) * torch.rand([1])).long().to(self.device)
            target_labels[counter] = l[t]

        return target_labels.long().to(self.device)

    def __call__(self, inputs, labels=None, *args, **kwargs):
        given_training = self.model.training
        self._change_model_mode(given_training)

        if self._normalization_applied is True:
            inputs = self.inverse_normalize(inputs)
            self._set_normalization_applied(False)

            adv_inputs = self.forward(inputs, labels, *args, **kwargs)
            # adv_inputs = self.to_type(adv_inputs, self.return_type)

            adv_inputs = self.normalize(adv_inputs)
            self._set_normalization_applied(True)
        else:
            adv_inputs = self.forward(inputs, labels, *args, **kwargs)
            # adv_inputs = self.to_type(adv_inputs, self.return_type)

        self._recover_model_mode(given_training)

        return adv_inputs

    def __repr__(self):
        info = self.__dict__.copy()

        del_keys = ["model", "attack", "supported_mode"]

        for key in info.keys():
            if key[0] == "_":
                del_keys.append(key)

        for key in del_keys:
            del info[key]

        info["attack_mode"] = self.attack_mode
        info["normalization_used"] = (
            True if self.normalization_used is not None else False
        )

        return (
            self.attack
            + "("
            + ", ".join("{}={}".format(key, val) for key, val in info.items())
            + ")"
        )

    def __setattr__(self, name, value):
        object.__setattr__(self, name, value)

        attacks = self.__dict__.get("_attacks")

        # Get all items in iterable items.
        def get_all_values(items, stack=[]):
            if items not in stack:
                stack.append(items)
                if isinstance(items, list) or isinstance(items, dict):
                    if isinstance(items, dict):
                        items = list(items.keys()) + list(items.values())
                    for item in items:
                        yield from get_all_values(item, stack)
                else:
                    if isinstance(items, Attack):
                        yield items
            else:
                if isinstance(items, Attack):
                    yield items

        for num, value in enumerate(get_all_values(value)):
            attacks[name + "." + str(num)] = value
            for subname, subvalue in value.__dict__.get("_attacks").items():
                attacks[name + "." + subname] = subvalue