Evolutionary based black-box adversarial attack.

art/attacks/evasion/attack_ea.py

@@ -0,0 +1,260 @@
+"""
+This module implements the black-box attack `attack_EA`.
+
+| Paper link: https://www.sciencedirect.com/science/article/pii/S1568494623004155
+"""
+from __future__ import absolute_import, division, print_function, unicode_literals
+import logging
+from typing import Optional, TYPE_CHECKING
+import random
+import sys
+import numpy as np
+from art.attacks.attack import EvasionAttack
+from art.estimators.estimator import BaseEstimator, NeuralNetworkMixin
+from art.estimators.classification.classifier import ClassifierMixin
+
+random.seed(0)
+
+if TYPE_CHECKING:
+    from art.utils import CLASSIFIER_TYPE
+
+logger = logging.getLogger(__name__)
+
+
+class AttackEA(EvasionAttack):
+    """
+    This class implements the black-box attack `attack_ea`.
+
+     Paper link: https://www.sciencedirect.com/science/article/pii/S1568494623004155
+    """
+
+    attack_params = EvasionAttack.attack_params + [
+        "max_iter",
+        "confidence",
+        "targeted",
+    ]
+
+    _estimator_requirements = (BaseEstimator, ClassifierMixin, NeuralNetworkMixin)  # ?????
+
+    def __init__(
+        self,
+        classifier: "CLASSIFIER_TYPE",
+        max_iter: int = 10000,
+        confidence: float = 0.51,
+        targeted: bool = False,
+    ):
+        """
+        Create an attack_EA  attack instance.
+
+        :param classifier: A trained classifier predicting probabilities and not logits.
+        :param max_iter: The maximum number of iterations.
+        :param confidence: Confidence of adversarial examples
+        :param targeted: perform targeted attack
+        """
+        super().__init__(estimator=classifier)
+
+        self.max_iter = max_iter
+        self.confidence = confidence
+        self.targeted = targeted
+        self.pop_size = 40
+        self.number_of_elites = 10
+
+    @staticmethod
+    def _get_fitness(probs: np.ndarray) -> np.ndarray:
+        """
+         It simply returns the CNN's probability for the images but different objective functions can be used here.
+        :param probs: an array of images' probabilities of selected CNN
+        :return: returns images' probabilities in an array (40,)
+        """
+        fitness = probs
+        return fitness
+
+    def _selection_untargeted(self, images: np.ndarray, fitness: np.ndarray):
+        """
+        Population will be divided into elite, middle_class, and didn't make it based on
+        images (individuals) fitness values. The images furthest from the ancestor category will be
+        closer to be in the elite.
+        :param images: the population of images in an array: size (pop_size, 224, 224, 3)
+        :param fitness: an array of images' propabilities of selected CNN
+        :return: returns a tuple of elite, middle_class images, fitness values of elites, index number of elites
+                in the population array, and random_keep images as numpy arrays.
+        """
+        idx_elite = fitness.argsort()[: self.number_of_elites]
+        half_pop_size = images.shape[0] / 2
+        idx_middle_class = fitness.argsort()[self.number_of_elites : int(half_pop_size)]
+        elite = images[idx_elite, :]
+        middle_class = images[idx_middle_class, :]
+
+        possible_idx = set(range(0, images.shape[0])) - set(idx_elite)
+        idx_keep = random.sample(possible_idx, int(images.shape[0] / 2 - self.number_of_elites))
+        random_keep = images[idx_keep]
+        return elite, middle_class, random_keep
+
+    def _selection_targeted(self, images: np.ndarray, fitness: np.ndarray):
+        """
+        Population will be divided into elite, middle_class, and didn't make it based on
+        images (individuals) fitness values. The images closest to the target category will be
+        closer to be in the elite.
+        :param images: the population of images in an array: size (pop_size, 224, 224, 3)
+        :param fitness: an array of images' probabilities of selected CNN
+        :return: returns elite, middle_class images, fitness values of elites, index number of elites
+                in the population array, and random_keep images as numpy arrays.
+        """
+        idx_elite = fitness.argsort()[-self.number_of_elites :]
+        half_pop_size = images.shape[0] / 2
+        idx_middle_class = fitness.argsort()[int(half_pop_size) : -self.number_of_elites]
+        elite = images[idx_elite, :][::-1]
+        middle_class = images[idx_middle_class, :]
+
+        possible_idx = set(range(0, images.shape[0])) - set(idx_elite)
+        idx_keep = random.sample(possible_idx, int(images.shape[0] / 2 - self.number_of_elites))
+        random_keep = images[idx_keep]
+        return elite, middle_class, random_keep
+
+    @staticmethod
+    def _get_no_of_pixels(im_size: int) -> int:
+        """
+        :param im_size: Original inputs' size, represented by an integer value.
+        :return: returns an integer that will be used to decide how many pixels will be mutated
+        in the image during the current generation.
+        """
+        u_factor = np.random.uniform(0.0, 1.0)
+        n = 60  # normally 60, the smaller n -> more pixels to mutate
+        res = (u_factor ** (1.0 / (n + 1))) * im_size
+        no_of_pixels = im_size - res
+        return no_of_pixels
+
+    @staticmethod
+    def _mutation(
+        _x: np.ndarray,
+        no_of_pixels: int,
+        mutation_group: np.ndarray,
+        percentage: float,
+        boundary_min: int,
+        boundary_max: int,
+    ) -> np.ndarray:
+        """
+        :param _x: An array with the original input to be attacked.
+        :param no_of_pixels: An integer determines the number of pixels to mutate in the original input for the current
+            generation.
+        :param mutation_group: An array with the individuals which will be mutated
+        :param percentage: A decimal number from 0 to 1 that represents the percentage of individuals in the mutation
+            group that will undergo mutation.
+        :param boundary_min: keep the pixel within [0, 255]
+        :param boundary_max: keep the pixel within [0, 255]
+        :return: An array of mutated individuals
+        """
+        mutated_group = mutation_group.copy()
+        no_of_individuals = len(mutated_group)  # 10
+        for individual in range(int(no_of_individuals * percentage)):
+            locations_x = np.random.randint(_x.shape[0], size=int(no_of_pixels))
+            locations_y = np.random.randint(_x.shape[1], size=int(no_of_pixels))
+            locations_z = np.random.randint(_x.shape[2], size=int(no_of_pixels))
+            new_values: list[int] = random.choices([-1, 1], k=int(no_of_pixels))
+            mutated_group[individual, locations_x, locations_y, locations_z] = (
+                mutated_group[individual, locations_x, locations_y, locations_z] - new_values
+            )
+        mutated_group = np.clip(mutated_group, boundary_min, boundary_max)
+        return mutated_group
+
+    @staticmethod
+    def _get_crossover_parents(crossover_group: np.ndarray) -> list:
+        size = crossover_group.shape[0]
+        no_of_parents = random.randrange(0, size, 2)
+        parents_idx = random.sample(range(0, size), no_of_parents)
+        return parents_idx
+
+    @staticmethod
+    def _crossover(_x: np.ndarray, crossover_group: np.ndarray, parents_idx: list) -> np.ndarray:
+        crossedover_group = crossover_group.copy()
+        for i in range(0, len(parents_idx), 2):
+            parent_index_1 = parents_idx[i]
+            parent_index_2 = parents_idx[i + 1]
+            # 15% of the image will be crossovered.
+            crossover_range = int(_x.shape[0] * 0.15)
+            size_x = np.random.randint(0, crossover_range)
+            start_x = np.random.randint(0, _x.shape[0] - size_x)
+            size_y = np.random.randint(0, crossover_range)
+            start_y = np.random.randint(0, _x.shape[1] - size_y)
+            _z = np.random.randint(_x.shape[2])
+            temp = crossedover_group[parent_index_1, start_x : start_x + size_x, start_y : start_y + size_y, _z]
+            crossedover_group[
+                parent_index_1, start_x : start_x + size_x, start_y : start_y + size_y, _z
+            ] = crossedover_group[parent_index_2, start_x : start_x + size_x, start_y : start_y + size_y, _z]
+            crossedover_group[parent_index_2, start_x : start_x + size_x, start_y : start_y + size_y, _z] = temp
+        return crossedover_group
+
+    def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> np.ndarray:
+        """
+        :param x: An array with the original inputs to be attacked.
+        :param y: An integer with the true or target labels.
+        :return: An array holding the adversarial examples.
+        """
+        boundary_min = 0
+        boundary_max = 255
+        x_temp = np.expand_dims(x, axis=0)
+        pred = self.estimator.predict(x_temp)
+        anc_indx = np.argmax(pred)
+        print("Before the image index is: " + str(anc_indx))
+
+        if self.targeted:
+            print("Target class index number is: ", y)
+        # pop_size * ancestor images are created
+        images = np.array([x] * self.pop_size).astype(int)
+        count = 0
+        while True:
+            preds = self.estimator.predict(images)  # predictions of 40 images
+            dom_indx = np.argmax(preds[int(np.argmax(preds) / 1000)])  # best image indx
+            best_adv = images[int(np.argmax(preds) / 1000)]  # best adversarial image so far
+            dom_cat_prop = preds[int(np.argmax(preds) / 1000)][dom_indx]  # label probability
+            sys.stdout.write(
+                f"\rgeneration: {count}/{self.max_iter} ______ index: {dom_indx} ____ with confidence: {dom_cat_prop}  "
+            )
+            # Terminate the algorithm if:
+            if count == self.max_iter:
+                # if algorithm can not create the adversarial image within "max_iter" stop the algorithm.
+                print("\nFailed to generate adversarial image within", self.max_iter, " generations")
+                break
+            if not self.targeted and dom_indx != anc_indx:
+                # if the attack is not targeted, the algorithm stops as soon as the image is classified in a
+                # category other than its true category.
+                print("\nAdversarial image is generated successfully in", count, "generations")
+                break
+            if self.targeted and dom_indx == y and dom_cat_prop > self.confidence:
+                # if the attack is targeted, the algorithm stops when the image is classified in the target category y.
+                print("\nAdversarial image is generated successfully in", count, " generations")
+                break
+
+            percentage_middle_class = 1
+            percentage_keep = 1
+            # Select population classes based on fitness
+            if self.targeted:
+                probs = np.array(preds)[:, y]
+                fitness = self._get_fitness(probs)
+                elite, middle_class, random_keep = self._selection_targeted(images, fitness)
+            else:
+                probs = np.array(preds)[:, anc_indx]
+                fitness = self._get_fitness(probs)
+                elite, middle_class, random_keep = self._selection_untargeted(images, fitness)
+            elite2 = elite.copy()
+            keep = np.concatenate((elite2, random_keep))
+            # Reproduce individuals by mutating Elite and Middle class---------
+            # mutate and crossover individuals
+            im_size = x.shape[0] * x.shape[1] * x.shape[2]
+            no_of_pixels = self._get_no_of_pixels(im_size)
+            mutated_middle_class = self._mutation(
+                x, no_of_pixels, middle_class, percentage_middle_class, boundary_min, boundary_max
+            )
+            mutated_keep_group1 = self._mutation(x, no_of_pixels, keep, percentage_keep, boundary_min, boundary_max)
+            mutated_keep_group2 = self._mutation(
+                x, no_of_pixels, mutated_keep_group1, percentage_keep, boundary_min, boundary_max
+            )
+            all_ = np.concatenate((mutated_middle_class, mutated_keep_group2))
+            parents_idx = self._get_crossover_parents(all_)
+            crossover_group = self._crossover(x, all_, parents_idx)
+
+            # Create new population
+            images = np.concatenate((elite, crossover_group))
+
+            count += 1
+        return best_adv