Merge pull request #655 from lyzsea/jsma

adversarial example attack method -- jsma

fluid/adversarial/advbox/attacks/saliency.py

+"""
+This module provide the attack method for JSMA's implement.
+"""
+from __future__ import division
+
+import logging
+import random
+import numpy as np
+
+from .base import Attack
+
+
+class SaliencyMapAttack(Attack):
+    """
+    Implements the Saliency Map Attack.
+    The Jacobian-based Saliency Map Approach (Papernot et al. 2016).
+    Paper link: https://arxiv.org/pdf/1511.07528.pdf
+    """
+
+    def _apply(self,
+               adversary,
+               max_iter=2000,
+               fast=True,
+               theta=0.1,
+               max_perturbations_per_pixel=7):
+        """
+        Apply the JSMA attack.
+        Args:
+            adversary(Adversary): The Adversary object.
+            max_iter(int): The max iterations.
+            fast(bool): Whether evaluate the pixel influence on sum of residual classes.
+            theta(float): Perturbation per pixel relative to [min, max] range.
+            max_perturbations_per_pixel(int): The max count of perturbation per pixel.
+        Return:
+            adversary: The Adversary object.
+        """
+        assert adversary is not None
+
+        if not adversary.is_targeted_attack or (adversary.target_label is None):
+            target_labels = self._generate_random_target(
+                adversary.original_label)
+        else:
+            target_labels = [adversary.target_label]
+
+        for target in target_labels:
+            original_image = adversary.original
+
+            # the mask defines the search domain
+            # each modified pixel with border value is set to zero in mask
+            mask = np.ones_like(original_image)
+
+            # count tracks how often each pixel was changed
+            counts = np.zeros_like(original_image)
+
+            labels = range(self.model.num_classes())
+            adv_img = original_image.copy()
+            min_, max_ = self.model.bounds()
+
+            for step in range(max_iter):
+                adv_img = np.clip(adv_img, min_, max_)
+                adv_label = np.argmax(self.model.predict(adv_img))
+                if adversary.try_accept_the_example(adv_img, adv_label):
+                    return adversary
+
+                # stop if mask is all zero
+                if not any(mask.flatten()):
+                    return adversary
+
+                logging.info('step = {}, original_label = {}, adv_label={}'.
+                             format(step, adversary.original_label, adv_label))
+
+                # get pixel location with highest influence on class
+                idx, p_sign = self._saliency_map(
+                    adv_img, target, labels, mask, fast=fast)
+
+                # apply perturbation
+                adv_img[idx] += -p_sign * theta * (max_ - min_)
+
+                # tracks number of updates for each pixel
+                counts[idx] += 1
+
+                # remove pixel from search domain if it hits the bound
+                if adv_img[idx] <= min_ or adv_img[idx] >= max_:
+                    mask[idx] = 0
+
+                # remove pixel if it was changed too often
+                if counts[idx] >= max_perturbations_per_pixel:
+                    mask[idx] = 0
+
+                adv_img = np.clip(adv_img, min_, max_)
+
+    def _generate_random_target(self, original_label):
+        """
+        Draw random target labels all of which are different and not the original label.
+        Args:
+            original_label(int): Original label.
+        Return:
+            target_labels(list): random target labels
+        """
+        num_random_target = 1
+        num_classes = self.model.num_classes()
+        assert num_random_target <= num_classes - 1
+
+        target_labels = random.sample(range(num_classes), num_random_target + 1)
+        target_labels = [t for t in target_labels if t != original_label]
+        target_labels = target_labels[:num_random_target]
+
+        return target_labels
+
+    def _saliency_map(self, image, target, labels, mask, fast=False):
+        """
+        Get pixel location with highest influence on class.
+        Args:
+            image(numpy.ndarray): Image with shape (height, width, channels).
+            target(int): The target label.
+            labels(int): The number of classes of the output label.
+            mask(list): Each modified pixel with border value is set to zero in mask.
+            fast(bool): Whether evaluate the pixel influence on sum of residual classes.
+        Return:
+            idx: The index of optimal pixel.
+            pix_sign: The direction of perturbation
+        """
+        # pixel influence on target class
+        alphas = self.model.gradient(image, target) * mask
+
+        # pixel influence on sum of residual classes(don't evaluate if fast == True)
+        if fast:
+            betas = -np.ones_like(alphas)
+        else:
+            betas = np.sum([
+                self.model.gradient(image, label) * mask - alphas
+                for label in labels
+            ], 0)
+
+        # compute saliency map (take into account both pos. & neg. perturbations)
+        sal_map = np.abs(alphas) * np.abs(betas) * np.sign(alphas * betas)
+
+        # find optimal pixel & direction of perturbation
+        idx = np.argmin(sal_map)
+        idx = np.unravel_index(idx, mask.shape)
+        pix_sign = np.sign(alphas)[idx]
+
+        return idx, pix_sign
+
+
+JSMA = SaliencyMapAttack

fluid/adversarial/mnist_tutorial_jsma.py

+"""
+FGSM demos on mnist using advbox tool.
+"""
+import matplotlib.pyplot as plt
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import numpy as np
+
+from advbox import Adversary
+from advbox.attacks.saliency import SaliencyMapAttack
+from advbox.models.paddle import PaddleModel
+
+
+def cnn_model(img):
+    """
+    Mnist cnn model
+    Args:
+        img(Varaible): the input image to be recognized
+    Returns:
+        Variable: the label prediction
+    """
+    # conv1 = fluid.nets.conv2d()
+    conv_pool_1 = fluid.nets.simple_img_conv_pool(
+        input=img,
+        num_filters=20,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    conv_pool_2 = fluid.nets.simple_img_conv_pool(
+        input=conv_pool_1,
+        num_filters=50,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
+    return logits
+
+
+def main():
+    """
+    Advbox demo which demonstrate how to use advbox.
+    """
+    IMG_NAME = 'img'
+    LABEL_NAME = 'label'
+
+    img = fluid.layers.data(name=IMG_NAME, shape=[1, 28, 28], dtype='float32')
+    # gradient should flow
+    img.stop_gradient = False
+    label = fluid.layers.data(name=LABEL_NAME, shape=[1], dtype='int64')
+    logits = cnn_model(img)
+    cost = fluid.layers.cross_entropy(input=logits, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+
+    BATCH_SIZE = 1
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=500),
+        batch_size=BATCH_SIZE)
+    feeder = fluid.DataFeeder(
+        feed_list=[IMG_NAME, LABEL_NAME],
+        place=place,
+        program=fluid.default_main_program())
+
+    fluid.io.load_params(
+        exe, "./mnist/", main_program=fluid.default_main_program())
+
+    # advbox demo
+    m = PaddleModel(fluid.default_main_program(), IMG_NAME, LABEL_NAME,
+                    logits.name, avg_cost.name, (-1, 1))
+    attack = SaliencyMapAttack(m)
+    total_num = 0
+    success_num = 0
+    for data in train_reader():
+        total_num += 1
+        # adversary.set_target(True, target_label=target_label)
+        jsma_attack = attack(Adversary(data[0][0], data[0][1]))
+        if jsma_attack is not None and jsma_attack.is_successful():
+            # plt.imshow(jsma_attack.target, cmap='Greys_r')
+            # plt.show()
+            success_num += 1
+            print('original_label=%d, adversary examples label =%d' %
+                  (data[0][1], jsma_attack.adversarial_label))
+            # np.save('adv_img', jsma_attack.adversarial_example)
+        print('total num = %d, success num = %d ' % (total_num, success_num))
+        if total_num == 100:
+            break
+
+
+if __name__ == '__main__':
+    main()