add fawkes-lite and fawkes-dev

fawkes_dev/README.md

@@ -0,0 +1,57 @@
+# Fawkes
+Code implementation of the paper "[Fawkes: Protecting Personal Privacy against Unauthorized Deep Learning Models](https://arxiv.org/pdf/2002.08327.pdf)", at *USENIX Security 2020*. 
+
+### BEFORE YOU RUN OUR CODE
+We appreciate your interest in our work and for trying out our code. We've noticed several cases where incorrect configuration leads to poor performances of protection. If you also observe low detection performance far away from what we presented in the paper, please feel free to open an issue in this repo or contact any of the authors directly. We are more than happy to help you debug your experiment and find out the correct configuration. 
+
+### ABOUT
+
+This repository contains code implementation of the paper "[Fawkes: Protecting Personal Privacy against Unauthorized Deep Learning Models](https://arxiv.org/pdf/2002.08327.pdf)", at *USENIX Security 2020*. 
+
+### DEPENDENCIES
+
+Our code is implemented and tested on Keras with TensorFlow backend. Following packages are used by our code.
+
+- `keras==2.3.1`
+- `numpy==1.18.4`
+- `tensorflow-gpu==1.13.1`
+
+Our code is tested on `Python 3.6.8`
+
+### HOWTO
+
+#### Download and Config Datasets
+The first step is to download several datasets for protection and target selection. 
+1. Download the following dataset to your local machine. After downloading the datasets, restructure it the same way as the FaceScrub dataset downloaded. 
+    - FaceScrub -- used for protection evaluation (link)
+    - VGGFace1 -- used for target select (link)
+    - VGGFace2 -- used for target select (link)
+    - WebFace -- used for target select (link)
+
+2. Config datasets
+open `fawkes/config.py` and update the `DATASETS` dictionary with the path to each dataset. Then run `python fawkes/config.py`. Every time the datasets are updated or moved, remember to rerun the command with the updated path. 
+
+3. Calculate embeddings using feature extractor. 
+Run `python3 fawkes/prepare_feature_extractor.py --candidate-datasets scrub vggface1 vggface2 webface`. This will calculate and cache the embeddings using the default feature extractor we provide. To use a customized feature extractor, please look at the Advance section at the end. 
+
+#### Generate Cloak for Images
+To generate cloak, run 
+`python3 fawkes/protection.py --gpu 0 --dataset scrub --feature-extractor webface_dense_robust_extract`
+For more information about the detailed parameters, please read `fawkes/protection.py`. 
+The code will output a directory in `results/` with `cloak_data.p` inside. You can check the cloaked images or inspect the changes in `this notebook`. 
+
+#### Evaluate Cloak Effectiveness
+To evaluate the cloak, run `python3 fawkes/eval_cloak.py --gpu 0 --cloak_data PATH-TO-RESULT-DIRECTORY --transfer_model vggface2_inception_extract`. 
+
+The code will print out the tracker model accuracy on uncloaked/original test images of the protected user, which should be close to 0. 
+
+
+### Citation
+```
+@inproceedings{shan2020fawkes,
+  title={Fawkes: Protecting Personal Privacy against Unauthorized Deep Learning Models},
+  author={Shan, Shawn and Wenger, Emily and Zhang, Jiayun and Li, Huiying and Zheng, Haitao and Zhao, Ben Y},
+  booktitle="Proc. of USENIX Security",
+  year={2020}
+}
+```

fawkes_dev/config.py

@@ -0,0 +1,40 @@
+import glob
+import json
+import os
+
+DATASETS = {
+    "scrub": "../data/scrub/",
+    "vggface1": "/mnt/data/sixiongshan/data/vggface/",
+    # "vggface2": "/mnt/data/sixiongshan/data/vggface2/",
+    "webface": "/mnt/data/sixiongshan/data/webface/",
+    # "youtubeface": "/mnt/data/sixiongshan/data/youtubeface/keras_flow_data/",
+}
+
+
+def main():
+    config = {}
+    for dataset in DATASETS.keys():
+        path = DATASETS[dataset]
+        if not os.path.exists(path):
+            print("Dataset path for {} does not exist, skipped".format(dataset))
+            continue
+        train_dir = os.path.join(path, "train")
+        test_dir = os.path.join(path, "test")
+        if not os.path.exists(train_dir):
+            print("Training dataset path for {} does not exist, skipped".format(dataset))
+            continue
+        num_classes = len(os.listdir(train_dir))
+        num_images = len(glob.glob(os.path.join(train_dir, "*/*")))
+        if num_images == 0 or num_classes == 0 or num_images == num_classes:
+            raise Exception("Dataset {} is not setup as detailed in README.".format(dataset))
+
+        config[dataset] = {"train_dir": train_dir, "test_dir": test_dir, "num_classes": num_classes,
+                           "num_images": num_images}
+        print("Successfully config {}".format(dataset))
+    j = json.dumps(config)
+    with open("config.json", "wb") as f:
+        f.write(j.encode())
+
+
+if __name__ == '__main__':
+    main()

fawkes_dev/differentiator.py

@@ -0,0 +1,431 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# @Date    : 2020-05-17
+# @Author  : Shawn Shan (shansixiong@cs.uchicago.edu)
+# @Link    : https://www.shawnshan.com/
+
+import datetime
+import time
+from decimal import Decimal
+
+import numpy as np
+import tensorflow as tf
+from utils import preprocess, reverse_preprocess
+
+
+class FawkesMaskGeneration:
+    # if the attack is trying to mimic a target image or a neuron vector
+    MIMIC_IMG = True
+    # number of iterations to perform gradient descent
+    MAX_ITERATIONS = 10000
+    # larger values converge faster to less accurate results
+    LEARNING_RATE = 1e-2
+    # the initial constant c to pick as a first guess
+    INITIAL_CONST = 1
+    # pixel intensity range
+    INTENSITY_RANGE = 'imagenet'
+    # threshold for distance
+    L_THRESHOLD = 0.03
+    # whether keep the final result or the best result
+    KEEP_FINAL = False
+    # max_val of image
+    MAX_VAL = 255
+    # The following variables are used by DSSIM, should keep as default
+    # filter size in SSIM
+    FILTER_SIZE = 11
+    # filter sigma in SSIM
+    FILTER_SIGMA = 1.5
+    # weights used in MS-SSIM
+    SCALE_WEIGHTS = None
+    MAXIMIZE = False
+    IMAGE_SHAPE = (224, 224, 3)
+    RATIO = 1.0
+    LIMIT_DIST = False
+
+    def __init__(self, sess, bottleneck_model_ls, mimic_img=MIMIC_IMG,
+                 batch_size=1, learning_rate=LEARNING_RATE,
+                 max_iterations=MAX_ITERATIONS, initial_const=INITIAL_CONST,
+                 intensity_range=INTENSITY_RANGE, l_threshold=L_THRESHOLD,
+                 max_val=MAX_VAL, keep_final=KEEP_FINAL, maximize=MAXIMIZE, image_shape=IMAGE_SHAPE,
+                 verbose=0, ratio=RATIO, limit_dist=LIMIT_DIST):
+
+        assert intensity_range in {'raw', 'imagenet', 'inception', 'mnist'}
+
+        # constant used for tanh transformation to avoid corner cases
+        self.tanh_constant = 2 - 1e-6
+        self.sess = sess
+        self.MIMIC_IMG = mimic_img
+        self.LEARNING_RATE = learning_rate
+        self.MAX_ITERATIONS = max_iterations
+        self.initial_const = initial_const
+        self.batch_size = batch_size
+        self.intensity_range = intensity_range
+        self.l_threshold = l_threshold
+        self.max_val = max_val
+        self.keep_final = keep_final
+        self.verbose = verbose
+        self.maximize = maximize
+        self.learning_rate = learning_rate
+        self.ratio = ratio
+        self.limit_dist = limit_dist
+        self.single_shape = list(image_shape)
+
+        self.input_shape = tuple([self.batch_size] + self.single_shape)
+
+        self.bottleneck_shape = tuple([self.batch_size] + self.single_shape)
+
+        # the variable we're going to optimize over
+        self.modifier = tf.Variable(np.zeros(self.input_shape, dtype=np.float32))
+
+        # target image in tanh space
+        if self.MIMIC_IMG:
+            self.timg_tanh = tf.Variable(np.zeros(self.input_shape), dtype=np.float32)
+        else:
+            self.bottleneck_t_raw = tf.Variable(np.zeros(self.bottleneck_shape), dtype=np.float32)
+        # source image in tanh space
+        self.simg_tanh = tf.Variable(np.zeros(self.input_shape), dtype=np.float32)
+
+        self.const = tf.Variable(np.ones(batch_size), dtype=np.float32)
+        self.mask = tf.Variable(np.ones((batch_size), dtype=np.bool))
+        self.weights = tf.Variable(np.ones(self.bottleneck_shape,
+                                           dtype=np.float32))
+
+        # and here's what we use to assign them
+        self.assign_modifier = tf.placeholder(tf.float32, self.input_shape)
+        if self.MIMIC_IMG:
+            self.assign_timg_tanh = tf.placeholder(
+                tf.float32, self.input_shape)
+        else:
+            self.assign_bottleneck_t_raw = tf.placeholder(
+                tf.float32, self.bottleneck_shape)
+        self.assign_simg_tanh = tf.placeholder(tf.float32, self.input_shape)
+        self.assign_const = tf.placeholder(tf.float32, (batch_size))
+        self.assign_mask = tf.placeholder(tf.bool, (batch_size))
+        self.assign_weights = tf.placeholder(tf.float32, self.bottleneck_shape)
+
+        # the resulting image, tanh'd to keep bounded from -0.5 to 0.5
+        # adversarial image in raw space
+        self.aimg_raw = (tf.tanh(self.modifier + self.simg_tanh) /
+                         self.tanh_constant +
+                         0.5) * 255.0
+        # source image in raw space
+        self.simg_raw = (tf.tanh(self.simg_tanh) /
+                         self.tanh_constant +
+                         0.5) * 255.0
+        if self.MIMIC_IMG:
+            # target image in raw space
+            self.timg_raw = (tf.tanh(self.timg_tanh) /
+                             self.tanh_constant +
+                             0.5) * 255.0
+
+        # convert source and adversarial image into input space
+        if self.intensity_range == 'imagenet':
+            mean = tf.constant(np.repeat([[[[103.939, 116.779, 123.68]]]], self.batch_size, axis=0), dtype=tf.float32,
+                               name='img_mean')
+            self.aimg_input = (self.aimg_raw[..., ::-1] - mean)
+            self.simg_input = (self.simg_raw[..., ::-1] - mean)
+            if self.MIMIC_IMG:
+                self.timg_input = (self.timg_raw[..., ::-1] - mean)
+
+        elif self.intensity_range == 'raw':
+            self.aimg_input = self.aimg_raw
+            self.simg_input = self.simg_raw
+            if self.MIMIC_IMG:
+                self.timg_input = self.timg_raw
+
+        def batch_gen_DSSIM(aimg_raw_split, simg_raw_split):
+            msssim_split = tf.image.ssim(aimg_raw_split, simg_raw_split, max_val=255.0)
+            dist = (1.0 - tf.stack(msssim_split)) / 2.0
+            return dist
+
+        # raw value of DSSIM distance
+        self.dist_raw = batch_gen_DSSIM(self.aimg_raw, self.simg_raw)
+        # distance value after applying threshold
+        self.dist = tf.maximum(self.dist_raw - self.l_threshold, 0.0)
+
+        self.dist_raw_sum = tf.reduce_sum(
+            tf.where(self.mask,
+                     self.dist_raw,
+                     tf.zeros_like(self.dist_raw)))
+        self.dist_sum = tf.reduce_sum(tf.where(self.mask, self.dist, tf.zeros_like(self.dist)))
+
+        def resize_tensor(input_tensor, model_input_shape):
+            if input_tensor.shape[1:] == model_input_shape or model_input_shape[1] is None:
+                return input_tensor
+            resized_tensor = tf.image.resize(input_tensor, model_input_shape[:2])
+            return resized_tensor
+
+        def calculate_direction(bottleneck_model, cur_timg_input, cur_simg_input):
+            target_features = bottleneck_model(cur_timg_input)
+            return target_features
+            # target_center = tf.reduce_mean(target_features, axis=0)
+            # original = bottleneck_model(cur_simg_input)
+            # original_center = tf.reduce_mean(original, axis=0)
+            # direction = target_center - original_center
+            # final_target = original + self.ratio * direction
+            # return final_target
+
+        self.bottlesim = 0.0
+        self.bottlesim_sum = 0.0
+        self.bottlesim_push = 0.0
+        for bottleneck_model in bottleneck_model_ls:
+            model_input_shape = bottleneck_model.input_shape[1:]
+            cur_aimg_input = resize_tensor(self.aimg_input, model_input_shape)
+
+            self.bottleneck_a = bottleneck_model(cur_aimg_input)
+            if self.MIMIC_IMG:
+                # cur_timg_input = resize_tensor(self.timg_input, model_input_shape)
+                # cur_simg_input = resize_tensor(self.simg_input, model_input_shape)
+                cur_timg_input = self.timg_input
+                cur_simg_input = self.simg_input
+                self.bottleneck_t = calculate_direction(bottleneck_model, cur_timg_input, cur_simg_input)
+                # self.bottleneck_t = bottleneck_model(cur_timg_input)
+            else:
+                self.bottleneck_t = self.bottleneck_t_raw
+
+            bottleneck_diff = self.bottleneck_t - self.bottleneck_a
+            scale_factor = tf.sqrt(tf.reduce_sum(tf.square(self.bottleneck_t), axis=1))
+
+            cur_bottlesim = tf.sqrt(tf.reduce_sum(tf.square(bottleneck_diff), axis=1))
+            cur_bottlesim = cur_bottlesim / scale_factor
+            cur_bottlesim_sum = tf.reduce_sum(cur_bottlesim)
+
+            self.bottlesim += cur_bottlesim
+
+            # self.bottlesim_push += cur_bottlesim_push_sum
+            self.bottlesim_sum += cur_bottlesim_sum
+
+        # sum up the losses
+        if self.maximize:
+            self.loss = self.const * tf.square(self.dist) - self.bottlesim
+        else:
+            self.loss = self.const * tf.square(self.dist) + self.bottlesim
+
+        self.loss_sum = tf.reduce_sum(tf.where(self.mask, self.loss, tf.zeros_like(self.loss)))
+
+        # Setup the Adadelta optimizer and keep track of variables
+        # we're creating
+        start_vars = set(x.name for x in tf.global_variables())
+        self.learning_rate_holder = tf.placeholder(tf.float32, shape=[])
+        # optimizer = tf.train.AdadeltaOptimizer(self.learning_rate_holder)
+        optimizer = tf.train.AdamOptimizer(self.learning_rate_holder)
+
+        self.train = optimizer.minimize(self.loss_sum,
+                                        var_list=[self.modifier])
+        end_vars = tf.global_variables()
+        new_vars = [x for x in end_vars if x.name not in start_vars]
+
+        # these are the variables to initialize when we run
+        self.setup = []
+        self.setup.append(self.modifier.assign(self.assign_modifier))
+        if self.MIMIC_IMG:
+            self.setup.append(self.timg_tanh.assign(self.assign_timg_tanh))
+        else:
+            self.setup.append(self.bottleneck_t_raw.assign(
+                self.assign_bottleneck_t_raw))
+        self.setup.append(self.simg_tanh.assign(self.assign_simg_tanh))
+        self.setup.append(self.const.assign(self.assign_const))
+        self.setup.append(self.mask.assign(self.assign_mask))
+        self.setup.append(self.weights.assign(self.assign_weights))
+
+        self.init = tf.variables_initializer(var_list=[self.modifier] + new_vars)
+
+        print('Attacker loaded')
+
+    def preprocess_arctanh(self, imgs):
+
+        imgs = reverse_preprocess(imgs, self.intensity_range)
+        imgs /= 255.0
+        imgs -= 0.5
+        imgs *= self.tanh_constant
+        tanh_imgs = np.arctanh(imgs)
+
+        return tanh_imgs
+
+    def clipping(self, imgs):
+
+        imgs = reverse_preprocess(imgs, self.intensity_range)
+        imgs = np.clip(imgs, 0, self.max_val)
+        imgs = np.rint(imgs)
+
+        imgs = preprocess(imgs, self.intensity_range)
+
+        return imgs
+
+    def attack(self, source_imgs, target_imgs, weights=None):
+
+        if weights is None:
+            weights = np.ones([source_imgs.shape[0]] +
+                              list(self.bottleneck_shape[1:]))
+
+        assert weights.shape[1:] == self.bottleneck_shape[1:]
+        assert source_imgs.shape[1:] == self.input_shape[1:]
+        assert source_imgs.shape[0] == weights.shape[0]
+        if self.MIMIC_IMG:
+            assert target_imgs.shape[1:] == self.input_shape[1:]
+            assert source_imgs.shape[0] == target_imgs.shape[0]
+        else:
+            assert target_imgs.shape[1:] == self.bottleneck_shape[1:]
+            assert source_imgs.shape[0] == target_imgs.shape[0]
+
+        start_time = time.time()
+
+        adv_imgs = []
+        print('%d batches in total'
+              % int(np.ceil(len(source_imgs) / self.batch_size)))
+
+        for idx in range(0, len(source_imgs), self.batch_size):
+            print('processing batch %d at %s' % (idx, datetime.datetime.now()))
+            adv_img = self.attack_batch(source_imgs[idx:idx + self.batch_size],
+                                        target_imgs[idx:idx + self.batch_size],
+                                        weights[idx:idx + self.batch_size])
+            adv_imgs.extend(adv_img)
+
+        elapsed_time = time.time() - start_time
+        print('attack cost %f s' % (elapsed_time))
+
+        return np.array(adv_imgs)
+
+    def attack_batch(self, source_imgs, target_imgs, weights):
+
+        """
+        Run the attack on a batch of images and labels.
+        """
+
+        LR = self.learning_rate
+        nb_imgs = source_imgs.shape[0]
+        mask = [True] * nb_imgs + [False] * (self.batch_size - nb_imgs)
+        mask = np.array(mask, dtype=np.bool)
+
+        source_imgs = np.array(source_imgs)
+        target_imgs = np.array(target_imgs)
+
+        # convert to tanh-space
+        simg_tanh = self.preprocess_arctanh(source_imgs)
+        if self.MIMIC_IMG:
+            timg_tanh = self.preprocess_arctanh(target_imgs)
+        else:
+            timg_tanh = target_imgs
+
+        CONST = np.ones(self.batch_size) * self.initial_const
+
+        self.sess.run(self.init)
+        simg_tanh_batch = np.zeros(self.input_shape)
+        if self.MIMIC_IMG:
+            timg_tanh_batch = np.zeros(self.input_shape)
+        else:
+            timg_tanh_batch = np.zeros(self.bottleneck_shape)
+        weights_batch = np.zeros(self.bottleneck_shape)
+        simg_tanh_batch[:nb_imgs] = simg_tanh[:nb_imgs]
+        timg_tanh_batch[:nb_imgs] = timg_tanh[:nb_imgs]
+        weights_batch[:nb_imgs] = weights[:nb_imgs]
+        modifier_batch = np.ones(self.input_shape) * 1e-6
+
+        # set the variables so that we don't have to send them over again
+        if self.MIMIC_IMG:
+            self.sess.run(self.setup,
+                          {self.assign_timg_tanh: timg_tanh_batch,
+                           self.assign_simg_tanh: simg_tanh_batch,
+                           self.assign_const: CONST,
+                           self.assign_mask: mask,
+                           self.assign_weights: weights_batch,
+                           self.assign_modifier: modifier_batch})
+        else:
+            # if directly mimicking a vector, use assign_bottleneck_t_raw
+            # in setup
+            self.sess.run(self.setup,
+                          {self.assign_bottleneck_t_raw: timg_tanh_batch,
+                           self.assign_simg_tanh: simg_tanh_batch,
+                           self.assign_const: CONST,
+                           self.assign_mask: mask,
+                           self.assign_weights: weights_batch,
+                           self.assign_modifier: modifier_batch})
+
+        best_bottlesim = [0] * nb_imgs if self.maximize else [np.inf] * nb_imgs
+        best_adv = np.zeros_like(source_imgs)
+
+        if self.verbose == 1:
+            loss_sum = float(self.sess.run(self.loss_sum))
+            dist_sum = float(self.sess.run(self.dist_sum))
+            thresh_over = (dist_sum / self.batch_size / self.l_threshold * 100)
+            dist_raw_sum = float(self.sess.run(self.dist_raw_sum))
+            bottlesim_sum = self.sess.run(self.bottlesim_sum)
+            print('START: Total loss: %.4E; perturb: %.6f (%.2f%% over, raw: %.6f); sim: %f'
+                  % (Decimal(loss_sum),
+                     dist_sum,
+                     thresh_over,
+                     dist_raw_sum,
+                     bottlesim_sum / nb_imgs))
+
+        try:
+            total_distance = [0] * nb_imgs
+
+            if self.limit_dist:
+                dist_raw_list, bottlesim_list, aimg_input_list = self.sess.run(
+                    [self.dist_raw,
+                     self.bottlesim,
+                     self.aimg_input])
+                for e, (dist_raw, bottlesim, aimg_input) in enumerate(
+                        zip(dist_raw_list, bottlesim_list, aimg_input_list)):
+                    if e >= nb_imgs:
+                        break
+                    total_distance[e] = bottlesim
+
+            for iteration in range(self.MAX_ITERATIONS):
+
+                self.sess.run([self.train], feed_dict={self.learning_rate_holder: LR})
+
+                dist_raw_list, bottlesim_list, aimg_input_list = self.sess.run(
+                    [self.dist_raw,
+                     self.bottlesim,
+                     self.aimg_input])
+                for e, (dist_raw, bottlesim, aimg_input) in enumerate(
+                        zip(dist_raw_list, bottlesim_list, aimg_input_list)):
+                    if e >= nb_imgs:
+                        break
+                    if (bottlesim < best_bottlesim[e] and bottlesim > total_distance[e] * 0.1 and (
+                            not self.maximize)) or (
+                            bottlesim > best_bottlesim[e] and self.maximize):
+                        best_bottlesim[e] = bottlesim
+                        best_adv[e] = aimg_input
+
+                if iteration != 0 and iteration % (self.MAX_ITERATIONS // 3) == 0:
+                    LR = LR / 2
+                    print("Learning Rate: ", LR)
+
+                if iteration % (self.MAX_ITERATIONS // 10) == 0:
+                    if self.verbose == 1:
+                        loss_sum = float(self.sess.run(self.loss_sum))
+                        dist_sum = float(self.sess.run(self.dist_sum))
+                        thresh_over = (dist_sum /
+                                       self.batch_size /
+                                       self.l_threshold *
+                                       100)
+                        dist_raw_sum = float(self.sess.run(self.dist_raw_sum))
+                        bottlesim_sum = self.sess.run(self.bottlesim_sum)
+                        print('ITER %4d: Total loss: %.4E; perturb: %.6f (%.2f%% over, raw: %.6f); sim: %f'
+                              % (iteration,
+                                 Decimal(loss_sum),
+                                 dist_sum,
+                                 thresh_over,
+                                 dist_raw_sum,
+                                 bottlesim_sum / nb_imgs))
+        except KeyboardInterrupt:
+            pass
+
+        if self.verbose == 1:
+            loss_sum = float(self.sess.run(self.loss_sum))
+            dist_sum = float(self.sess.run(self.dist_sum))
+            thresh_over = (dist_sum / self.batch_size / self.l_threshold * 100)
+            dist_raw_sum = float(self.sess.run(self.dist_raw_sum))
+            bottlesim_sum = float(self.sess.run(self.bottlesim_sum))
+            print('END:       Total loss: %.4E; perturb: %.6f (%.2f%% over, raw: %.6f); sim: %f'
+                  % (Decimal(loss_sum),
+                     dist_sum,
+                     thresh_over,
+                     dist_raw_sum,
+                     bottlesim_sum / nb_imgs))
+
+        best_adv = self.clipping(best_adv[:nb_imgs])
+
+        return best_adv

fawkes_dev/eval_cloak.py

@@ -0,0 +1,157 @@
+import sys
+
+sys.path.append("/home/shansixioing/tools/")
+sys.path.append("/home/shansixioing/cloak/")
+
+import argparse
+from tensorflow import set_random_seed
+from utils import init_gpu, load_extractor, load_victim_model, dump_dictionary_as_json
+import os
+import numpy as np
+import random
+import pickle
+import re
+from keras.preprocessing import image
+from keras.utils import to_categorical
+from keras.applications.vgg16 import preprocess_input
+
+# import locale
+#
+# loc = locale.getlocale()
+# locale.setlocale(locale.LC_ALL, loc)
+
+
+def select_samples(data_dir):
+    all_data_path = []
+    for cls in os.listdir(data_dir):
+        cls_dir = os.path.join(data_dir, cls)
+        for data_path in os.listdir(cls_dir):
+            all_data_path.append(os.path.join(cls_dir, data_path))
+
+    return all_data_path
+
+
+def generator_wrap(cloak_data, n_classes, test=False, validation_split=0.1):
+    if test:
+        all_data_path = select_samples(cloak_data.test_data_dir)
+    else:
+        all_data_path = select_samples(cloak_data.train_data_dir)
+    split = int(len(cloak_data.cloaked_protect_train_X) * (1 - validation_split))
+    cloaked_train_X = cloak_data.cloaked_protect_train_X[:split]
+    np.random.seed(12345)
+
+    # all_vals = list(cloak_data.path2idx.items())
+
+    while True:
+        batch_X = []
+        batch_Y = []
+        cur_batch_path = np.random.choice(all_data_path, args.batch_size)
+        for p in cur_batch_path:
+            # p = p.encode("utf-8").decode("ascii", 'ignore')
+            cur_y = cloak_data.path2idx[p]
+            # protect class and sybil class do not need to appear in test dataset
+            if test and (re.search(cloak_data.protect_class, p)):
+                continue
+            # protect class images in train dataset
+            elif p in cloak_data.protect_class_path:
+                cur_x = random.choice(cloaked_train_X)
+            else:
+                im = image.load_img(p, target_size=cloak_data.img_shape)
+                im = image.img_to_array(im)
+                cur_x = preprocess_input(im)
+            batch_X.append(cur_x)
+            batch_Y.append(cur_y)
+
+        batch_X = np.array(batch_X)
+        batch_Y = to_categorical(np.array(batch_Y), num_classes=n_classes)
+
+        yield batch_X, batch_Y
+
+
+def eval_uncloaked_test_data(cloak_data, n_classes):
+    original_label = cloak_data.path2idx[list(cloak_data.protect_class_path)[0]]
+    protect_test_X = cloak_data.protect_test_X
+    original_Y = [original_label] * len(protect_test_X)
+    original_Y = to_categorical(original_Y, n_classes)
+    return protect_test_X, original_Y
+
+
+def eval_cloaked_test_data(cloak_data, n_classes, validation_split=0.1):
+    split = int(len(cloak_data.cloaked_protect_train_X) * (1 - validation_split))
+    cloaked_test_X = cloak_data.cloaked_protect_train_X[split:]
+    original_label = cloak_data.path2idx[list(cloak_data.protect_class_path)[0]]
+    original_Y = [original_label] * len(cloaked_test_X)
+    original_Y = to_categorical(original_Y, n_classes)
+    return cloaked_test_X, original_Y
+
+
+def main():
+    init_gpu(args.gpu)
+
+    if args.dataset == 'pubfig':
+        N_CLASSES = 65
+        CLOAK_DIR = args.cloak_data
+    elif args.dataset == 'scrub':
+        N_CLASSES = 530
+        CLOAK_DIR = args.cloak_data
+    else:
+        raise ValueError
+
+    CLOAK_DIR = os.path.join("../results", CLOAK_DIR)
+    RES = pickle.load(open(os.path.join(CLOAK_DIR, "cloak_data.p"), 'rb'))
+
+    print("Build attacker's model")
+    cloak_data = RES['cloak_data']
+    EVAL_RES = {}
+    train_generator = generator_wrap(cloak_data, n_classes=N_CLASSES,
+                                     validation_split=args.validation_split)
+    test_generator = generator_wrap(cloak_data, test=True, n_classes=N_CLASSES,
+                                    validation_split=args.validation_split)
+
+    EVAL_RES['transfer_model'] = args.transfer_model
+
+    base_model = load_extractor(args.transfer_model)
+    model = load_victim_model(teacher_model=base_model, number_classes=N_CLASSES)
+
+    original_X, original_Y = eval_uncloaked_test_data(cloak_data, N_CLASSES)
+    cloaked_test_X, cloaked_test_Y = eval_cloaked_test_data(cloak_data, N_CLASSES,
+                                                            validation_split=args.validation_split)
+
+    try:
+        model.fit_generator(train_generator, steps_per_epoch=cloak_data.number_samples // 32,
+                            validation_data=(original_X, original_Y), epochs=args.n_epochs, verbose=2,
+                            use_multiprocessing=False, workers=1)
+    except KeyboardInterrupt:
+        pass
+
+    _, acc_original = model.evaluate(original_X, original_Y, verbose=0)
+    print("Accuracy on uncloaked/original images TEST: {:.4f}".format(acc_original))
+    EVAL_RES['acc_original'] = acc_original
+
+    _, other_acc = model.evaluate_generator(test_generator, verbose=0, steps=50)
+    print("Accuracy on other classes {:.4f}".format(other_acc))
+    EVAL_RES['other_acc'] = other_acc
+    dump_dictionary_as_json(EVAL_RES, os.path.join(CLOAK_DIR, "eval_seed{}.json".format(args.seed_idx)))
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--gpu', type=str,
+                        help='GPU id', default='0')
+    parser.add_argument('--dataset', type=str,
+                        help='name of dataset', default='scrub')
+    parser.add_argument('--cloak_data', type=str,
+                        help='name of the cloak result directory',
+                        default='scrub_webface_dense_robust_extract_protectPatrick_Dempsey')
+    parser.add_argument('--transfer_model', type=str,
+                        help='the feature extractor used for tracker model training. It can be the same or not same as the user\'s', default='vggface2_inception_extract')
+    parser.add_argument('--batch_size', type=int, default=32)
+    parser.add_argument('--validation_split', type=float, default=0.1)
+    parser.add_argument('--n_epochs', type=int, default=5)
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    args = parse_arguments(sys.argv[1:])
+    main()

fawkes_dev/prepare_feature_extractor.py

@@ -0,0 +1,64 @@
+import argparse
+import os
+import pickle
+import random
+import sys
+
+import numpy as np
+from keras.applications.vgg16 import preprocess_input
+from keras.preprocessing import image
+from utils import load_extractor, get_dataset_path
+
+
+def load_sample_dir(path, sample=10):
+    x_ls = []
+    image_paths = list(os.listdir(path))
+    random.shuffle(image_paths)
+    for i, file in enumerate(image_paths):
+        if i > sample:
+            break
+        cur_path = os.path.join(path, file)
+        im = image.load_img(cur_path, target_size=(224, 224))
+        im = image.img_to_array(im)
+        x_ls.append(im)
+    raw_x = np.array(x_ls)
+    return preprocess_input(raw_x)
+
+
+def normalize(x):
+    return x / np.linalg.norm(x)
+
+
+def main():
+    extractor = load_extractor(args.feature_extractor)
+    path2emb = {}
+    for target_dataset in args.candidate_datasets:
+        target_dataset_path, _, _, _ = get_dataset_path(target_dataset)
+        for target_class in os.listdir(target_dataset_path):
+            target_class_path = os.path.join(target_dataset_path, target_class)
+            target_X = load_sample_dir(target_class_path)
+            cur_feature = extractor.predict(target_X)
+            cur_feature = np.mean(cur_feature, axis=0)
+            path2emb[target_class_path] = cur_feature
+
+    for k, v in path2emb.items():
+        path2emb[k] = normalize(v)
+
+    pickle.dump(path2emb, open("../feature_extractors/embeddings/{}_emb_norm.p".format(args.feature_extractor), "wb"))
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--gpu', type=str,
+                        help='GPU id', default='0')
+    parser.add_argument('--candidate-datasets', nargs='+',
+                        help='path candidate datasets')
+    parser.add_argument('--feature-extractor', type=str,
+                        help="name of the feature extractor used for optimization",
+                        default="webface_dense_robust_extract")
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    args = parse_arguments(sys.argv[1:])
+    main()

fawkes_dev/protection.py

@@ -0,0 +1,95 @@
+import argparse
+import os
+import pickle
+import random
+import sys
+
+import numpy as np
+from differentiator import FawkesMaskGeneration
+from tensorflow import set_random_seed
+from utils import load_extractor, CloakData, init_gpu
+
+random.seed(12243)
+np.random.seed(122412)
+set_random_seed(12242)
+
+NUM_IMG_PROTECTED = 32  # Number of images used to optimize the target class
+BATCH_SIZE = 32
+
+MAX_ITER = 1000
+
+
+def diff_protected_data(sess, feature_extractors_ls, image_X, number_protect, target_X=None, th=0.01):
+    image_X = image_X[:number_protect]
+    differentiator = FawkesMaskGeneration(sess, feature_extractors_ls,
+                                          batch_size=BATCH_SIZE,
+                                          mimic_img=True,
+                                          intensity_range='imagenet',
+                                          initial_const=args.sd,
+                                          learning_rate=args.lr,
+                                          max_iterations=MAX_ITER,
+                                          l_threshold=th,
+                                          verbose=1, maximize=False, keep_final=False, image_shape=image_X.shape[1:])
+
+    if len(target_X) < len(image_X):
+        target_X = np.concatenate([target_X, target_X, target_X])
+    target_X = target_X[:len(image_X)]
+    cloaked_image_X = differentiator.attack(image_X, target_X)
+    return cloaked_image_X
+
+
+def perform_defense():
+    RES = {}
+    sess = init_gpu(args.gpu)
+
+    FEATURE_EXTRACTORS = [args.feature_extractor]
+    RES_DIR = '../results/'
+
+    RES['num_img_protected'] = NUM_IMG_PROTECTED
+    RES['extractors'] = FEATURE_EXTRACTORS
+    num_protect = NUM_IMG_PROTECTED
+
+    print("Loading {} for optimization".format(args.feature_extractor))
+    feature_extractors_ls = [load_extractor(name) for name in FEATURE_EXTRACTORS]
+    protect_class = args.protect_class
+
+    cloak_data = CloakData(args.dataset, protect_class=protect_class)
+    RES_FILE_NAME = "{}_{}_protect{}".format(args.dataset, args.feature_extractor, cloak_data.protect_class)
+    RES_FILE_NAME = os.path.join(RES_DIR, RES_FILE_NAME)
+    print("Protect Class: ", cloak_data.protect_class)
+
+    cloak_data.target_path, cloak_data.target_data = cloak_data.select_target_label(feature_extractors_ls,
+                                                                                    FEATURE_EXTRACTORS)
+
+    os.makedirs(RES_DIR, exist_ok=True)
+    os.makedirs(RES_FILE_NAME, exist_ok=True)
+
+    cloak_image_X = diff_protected_data(sess, feature_extractors_ls, cloak_data.protect_train_X,
+                                        number_protect=num_protect,
+                                        target_X=cloak_data.target_data, th=args.th)
+
+    cloak_data.cloaked_protect_train_X = cloak_image_X
+    RES['cloak_data'] = cloak_data
+    pickle.dump(RES, open(os.path.join(RES_FILE_NAME, 'cloak_data.p'), "wb"))
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--gpu', type=str,
+                        help='GPU id', default='0')
+    parser.add_argument('--dataset', type=str,
+                        help='name of dataset', default='scrub')
+    parser.add_argument('--feature-extractor', type=str,
+                        help="name of the feature extractor used for optimization",
+                        default="webface_dense_robust_extract")
+    parser.add_argument('--th', type=float, default=0.007)
+    parser.add_argument('--sd', type=int, default=1e5)
+    parser.add_argument('--protect_class', type=str, default=None)
+    parser.add_argument('--lr', type=float, default=0.1)
+
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    args = parse_arguments(sys.argv[1:])
+    perform_defense()

fawkes_dev/utils.py

@@ -0,0 +1,372 @@
+import json
+import os
+import pickle
+import random
+
+import keras
+import keras.backend as K
+import numpy as np
+import tensorflow as tf
+from keras.applications.vgg16 import preprocess_input
+from keras.layers import Dense, Activation
+from keras.models import Model
+from keras.preprocessing import image
+from keras.utils import to_categorical
+from sklearn.metrics import pairwise_distances
+
+
+def clip_img(X, preprocessing='raw'):
+    X = reverse_preprocess(X, preprocessing)
+    X = np.clip(X, 0.0, 255.0)
+    X = preprocess(X, preprocessing)
+    return X
+
+
+def dump_dictionary_as_json(dict, outfile):
+    j = json.dumps(dict)
+    with open(outfile, "wb") as f:
+        f.write(j.encode())
+
+
+def fix_gpu_memory(mem_fraction=1):
+    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
+    gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=mem_fraction)
+    tf_config = tf.ConfigProto(gpu_options=gpu_options)
+    tf_config.gpu_options.allow_growth = True
+    tf_config.log_device_placement = False
+    init_op = tf.global_variables_initializer()
+    sess = tf.Session(config=tf_config)
+    sess.run(init_op)
+    K.set_session(sess)
+    return sess
+
+
+def load_victim_model(number_classes, teacher_model=None, end2end=False):
+    for l in teacher_model.layers:
+        l.trainable = end2end
+    x = teacher_model.layers[-1].output
+
+    x = Dense(number_classes)(x)
+    x = Activation('softmax', name="act")(x)
+    model = Model(teacher_model.input, x)
+    opt = keras.optimizers.Adadelta()
+    model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
+    return model
+
+
+def init_gpu(gpu_index, force=False):
+    if isinstance(gpu_index, list):
+        gpu_num = ','.join([str(i) for i in gpu_index])
+    else:
+        gpu_num = str(gpu_index)
+    if "CUDA_VISIBLE_DEVICES" in os.environ and os.environ["CUDA_VISIBLE_DEVICES"] and not force:
+        print('GPU already initiated')
+        return
+    os.environ["CUDA_VISIBLE_DEVICES"] = gpu_num
+    sess = fix_gpu_memory()
+    return sess
+
+
+def preprocess(X, method):
+    assert method in {'raw', 'imagenet', 'inception', 'mnist'}
+
+    if method is 'raw':
+        pass
+    elif method is 'imagenet':
+        X = imagenet_preprocessing(X)
+    else:
+        raise Exception('unknown method %s' % method)
+
+    return X
+
+
+def reverse_preprocess(X, method):
+    assert method in {'raw', 'imagenet', 'inception', 'mnist'}
+
+    if method is 'raw':
+        pass
+    elif method is 'imagenet':
+        X = imagenet_reverse_preprocessing(X)
+    else:
+        raise Exception('unknown method %s' % method)
+
+    return X
+
+
+def imagenet_preprocessing(x, data_format=None):
+    if data_format is None:
+        data_format = K.image_data_format()
+    assert data_format in ('channels_last', 'channels_first')
+
+    x = np.array(x)
+    if data_format == 'channels_first':
+        # 'RGB'->'BGR'
+        if x.ndim == 3:
+            x = x[::-1, ...]
+        else:
+            x = x[:, ::-1, ...]
+    else:
+        # 'RGB'->'BGR'
+        x = x[..., ::-1]
+
+    mean = [103.939, 116.779, 123.68]
+    std = None
+
+    # Zero-center by mean pixel
+    if data_format == 'channels_first':
+        if x.ndim == 3:
+            x[0, :, :] -= mean[0]
+            x[1, :, :] -= mean[1]
+            x[2, :, :] -= mean[2]
+            if std is not None:
+                x[0, :, :] /= std[0]
+                x[1, :, :] /= std[1]
+                x[2, :, :] /= std[2]
+        else:
+            x[:, 0, :, :] -= mean[0]
+            x[:, 1, :, :] -= mean[1]
+            x[:, 2, :, :] -= mean[2]
+            if std is not None:
+                x[:, 0, :, :] /= std[0]
+                x[:, 1, :, :] /= std[1]
+                x[:, 2, :, :] /= std[2]
+    else:
+        x[..., 0] -= mean[0]
+        x[..., 1] -= mean[1]
+        x[..., 2] -= mean[2]
+        if std is not None:
+            x[..., 0] /= std[0]
+            x[..., 1] /= std[1]
+            x[..., 2] /= std[2]
+
+    return x
+
+
+def imagenet_reverse_preprocessing(x, data_format=None):
+    import keras.backend as K
+    x = np.array(x)
+    if data_format is None:
+        data_format = K.image_data_format()
+    assert data_format in ('channels_last', 'channels_first')
+
+    if data_format == 'channels_first':
+        if x.ndim == 3:
+            # Zero-center by mean pixel
+            x[0, :, :] += 103.939
+            x[1, :, :] += 116.779
+            x[2, :, :] += 123.68
+            # 'BGR'->'RGB'
+            x = x[::-1, :, :]
+        else:
+            x[:, 0, :, :] += 103.939
+            x[:, 1, :, :] += 116.779
+            x[:, 2, :, :] += 123.68
+            x = x[:, ::-1, :, :]
+    else:
+        # Zero-center by mean pixel
+        x[..., 0] += 103.939
+        x[..., 1] += 116.779
+        x[..., 2] += 123.68
+        # 'BGR'->'RGB'
+        x = x[..., ::-1]
+    return x
+
+
+def build_bottleneck_model(model, cut_off):
+    bottleneck_model = Model(model.input, model.get_layer(cut_off).output)
+    bottleneck_model.compile(loss='categorical_crossentropy',
+                             optimizer='adam',
+                             metrics=['accuracy'])
+    return bottleneck_model
+
+
+def load_extractor(name):
+    model = keras.models.load_model("../feature_extractors/{}.h5".format(name))
+    if hasattr(model.layers[-1], "activation") and model.layers[-1].activation == "softmax":
+        raise Exception(
+            "Given extractor's last layer is softmax, need to remove the top layers to make it into a feature extractor")
+    # if "extract" in name.split("/")[-1]:
+    #     pass
+    # else:
+    #     print("Convert a model to a feature extractor")
+    #     model = build_bottleneck_model(model, model.layers[layer_idx].name)
+    #     model.save(name + "extract")
+    #     model = keras.models.load_model(name + "extract")
+    return model
+
+
+def get_dataset_path(dataset):
+    if not os.path.exists("config.json"):
+        raise Exception("Please config the datasets before running protection code. See more in README and config.py.")
+
+    config = json.load(open("config.json", 'r'))
+    if dataset not in config:
+        raise Exception(
+            "Dataset {} does not exist, please download to data/ and add the path to this function... Abort".format(
+                dataset))
+    return config[dataset]['train_dir'], config[dataset]['test_dir'], config[dataset]['num_classes'], config[dataset][
+        'num_images']
+
+
+def normalize(x):
+    return x / np.linalg.norm(x, axis=1, keepdims=True)
+
+
+class CloakData(object):
+    def __init__(self, dataset, img_shape=(224, 224), protect_class=None):
+        self.dataset = dataset
+        self.img_shape = img_shape
+
+        self.train_data_dir, self.test_data_dir, self.number_classes, self.number_samples = get_dataset_path(dataset)
+        self.all_labels = sorted(list(os.listdir(self.train_data_dir)))
+        if protect_class:
+            self.protect_class = protect_class
+        else:
+            self.protect_class = random.choice(self.all_labels)
+
+        self.sybil_class = random.choice([l for l in self.all_labels if l != self.protect_class])
+        self.protect_train_X, self.protect_test_X = self.load_label_data(self.protect_class)
+        self.sybil_train_X, self.sybil_test_X = self.load_label_data(self.sybil_class)
+
+        self.cloaked_protect_train_X = None
+        self.cloaked_sybil_train_X = None
+
+        self.label2path_train, self.label2path_test, self.path2idx = self.build_data_mapping()
+        self.all_training_path = self.get_all_data_path(self.label2path_train)
+        self.all_test_path = self.get_all_data_path(self.label2path_test)
+        self.protect_class_path = self.get_class_image_files(os.path.join(self.train_data_dir, self.protect_class))
+        self.sybil_class_path = self.get_class_image_files(os.path.join(self.train_data_dir, self.sybil_class))
+
+        print("Find {} protect images".format(len(self.protect_class_path)))
+
+    def get_class_image_files(self, path):
+        return [os.path.join(path, f) for f in os.listdir(path)]
+
+    def extractor_ls_predict(self, feature_extractors_ls, X):
+        feature_ls = []
+        for extractor in feature_extractors_ls:
+            cur_features = extractor.predict(X)
+            feature_ls.append(cur_features)
+        concated_feature_ls = np.concatenate(feature_ls, axis=1)
+        concated_feature_ls = normalize(concated_feature_ls)
+        return concated_feature_ls
+
+    def load_embeddings(self, feature_extractors_names):
+        dictionaries = []
+        for extractor_name in feature_extractors_names:
+            path2emb = pickle.load(open("../feature_extractors/embeddings/{}_emb_norm.p".format(extractor_name), "rb"))
+            dictionaries.append(path2emb)
+
+        merge_dict = {}
+        for k in dictionaries[0].keys():
+            cur_emb = [dic[k] for dic in dictionaries]
+            merge_dict[k] = np.concatenate(cur_emb)
+        return merge_dict
+
+    def select_target_label(self, feature_extractors_ls, feature_extractors_names, metric='l2'):
+        original_feature_x = self.extractor_ls_predict(feature_extractors_ls, self.protect_train_X)
+
+        path2emb = self.load_embeddings(feature_extractors_names)
+        items = list(path2emb.items())
+        paths = [p[0] for p in items]
+        embs = [p[1] for p in items]
+        embs = np.array(embs)
+
+        pair_dist = pairwise_distances(original_feature_x, embs, metric)
+        max_sum = np.min(pair_dist, axis=0)
+        sorted_idx = np.argsort(max_sum)[::-1]
+
+        highest_num = 0
+        paired_target_X = None
+        final_target_class_path = None
+        for idx in sorted_idx[:5]:
+            target_class_path = paths[idx]
+            cur_target_X = self.load_dir(target_class_path)
+            cur_target_X = np.concatenate([cur_target_X, cur_target_X, cur_target_X])
+            cur_tot_sum, cur_paired_target_X = self.calculate_dist_score(self.protect_train_X, cur_target_X,
+                                                                         feature_extractors_ls,
+                                                                         metric=metric)
+            if cur_tot_sum > highest_num:
+                highest_num = cur_tot_sum
+                paired_target_X = cur_paired_target_X
+                final_target_class_path = target_class_path
+
+        np.random.shuffle(paired_target_X)
+        return final_target_class_path, paired_target_X
+
+    def calculate_dist_score(self, a, b, feature_extractors_ls, metric='l2'):
+        features1 = self.extractor_ls_predict(feature_extractors_ls, a)
+        features2 = self.extractor_ls_predict(feature_extractors_ls, b)
+
+        pair_cos = pairwise_distances(features1, features2, metric)
+        max_sum = np.min(pair_cos, axis=0)
+        max_sum_arg = np.argsort(max_sum)[::-1]
+        max_sum_arg = max_sum_arg[:len(a)]
+        max_sum = [max_sum[i] for i in max_sum_arg]
+        paired_target_X = [b[j] for j in max_sum_arg]
+        paired_target_X = np.array(paired_target_X)
+        return np.min(max_sum), paired_target_X
+
+    def get_all_data_path(self, label2path):
+        all_paths = []
+        for k, v in label2path.items():
+            cur_all_paths = [os.path.join(k, cur_p) for cur_p in v]
+            all_paths.extend(cur_all_paths)
+        return all_paths
+
+    def load_label_data(self, label):
+        train_label_path = os.path.join(self.train_data_dir, label)
+        test_label_path = os.path.join(self.test_data_dir, label)
+        train_X = self.load_dir(train_label_path)
+        test_X = self.load_dir(test_label_path)
+        return train_X, test_X
+
+    def load_dir(self, path):
+        assert os.path.exists(path)
+        x_ls = []
+        for file in os.listdir(path):
+            cur_path = os.path.join(path, file)
+            im = image.load_img(cur_path, target_size=self.img_shape)
+            im = image.img_to_array(im)
+            x_ls.append(im)
+        raw_x = np.array(x_ls)
+        return preprocess_input(raw_x)
+
+    def build_data_mapping(self):
+        label2path_train = {}
+        label2path_test = {}
+        idx = 0
+        path2idx = {}
+        for label_name in self.all_labels:
+            full_path_train = os.path.join(self.train_data_dir, label_name)
+            full_path_test = os.path.join(self.test_data_dir, label_name)
+            label2path_train[full_path_train] = list(os.listdir(full_path_train))
+            label2path_test[full_path_test] = list(os.listdir(full_path_test))
+            for img_file in os.listdir(full_path_train):
+                path2idx[os.path.join(full_path_train, img_file)] = idx
+            for img_file in os.listdir(full_path_test):
+                path2idx[os.path.join(full_path_test, img_file)] = idx
+            idx += 1
+        return label2path_train, label2path_test, path2idx
+
+    def generate_data_post_cloak(self, sybil=False):
+        assert self.cloaked_protect_train_X is not None
+        while True:
+            batch_X = []
+            batch_Y = []
+            cur_batch_path = random.sample(self.all_training_path, 32)
+            for p in cur_batch_path:
+                cur_y = self.path2idx[p]
+                if p in self.protect_class_path:
+                    cur_x = random.choice(self.cloaked_protect_train_X)
+                elif sybil and (p in self.sybil_class):
+                    cur_x = random.choice(self.cloaked_sybil_train_X)
+                else:
+                    im = image.load_img(p, target_size=self.img_shape)
+                    im = image.img_to_array(im)
+                    cur_x = preprocess_input(im)
+                batch_X.append(cur_x)
+                batch_Y.append(cur_y)
+            batch_X = np.array(batch_X)
+            batch_Y = to_categorical(np.array(batch_Y), num_classes=self.number_classes)
+            yield batch_X, batch_Y