fgsm.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "os.environ['KERAS_BACKEND'] = 'tensorflow'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "from matplotlib import pyplot as plt\n",
    "import numpy as np\n",
    "from keras.models import load_model\n",
    "from keras import backend as K\n",
    "import keras\n",
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_2 (InputLayer)         (None, 32, 32, 3)         0         \n",
      "_________________________________________________________________\n",
      "flatten_2 (Flatten)          (None, 3072)              0         \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 2048)              6293504   \n",
      "_________________________________________________________________\n",
      "dense_7 (Dense)              (None, 2048)              4196352   \n",
      "_________________________________________________________________\n",
      "dense_8 (Dense)              (None, 10)                20490     \n",
      "=================================================================\n",
      "Total params: 10,510,346\n",
      "Trainable params: 10,510,346\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = load_model(\"./saved_model/model.h5\")\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.datasets import cifar10\n",
    "(x_train, y_train), (x_test, y_test) = cifar10.load_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_train = x_train/255.0\n",
    "x_test = x_test/255.0\n",
    "mean=[0.485, 0.456, 0.406]\n",
    "std=[0.229, 0.224, 0.225]\n",
    "x_train = (x_train - mean)/std\n",
    "x_test = (x_test - mean)/std"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test accuracy: 0.5076\n"
     ]
    }
   ],
   "source": [
    "y_pred = model.predict(x_test, batch_size=64)\n",
    "y_pred = np.argmax(np.squeeze(y_pred), axis=1)\n",
    "y_test = np.squeeze(y_test)\n",
    "acc = sum(y_pred == y_test)/len(y_test)\n",
    "print(\"Test accuracy: \" + str(acc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "class AdversarialAttack(object):\n",
    "    def __init__(self, model, loss_func, epsilon=0.0):\n",
    "        self._loss_func = loss_func\n",
    "        self._epsilon = epsilon\n",
    "        self._model = keras.models.clone_model(model)\n",
    "\n",
    "        self._model.trainable = False\n",
    "        self._model.compile(optimizer=model.optimizer, loss=self._loss_func)\n",
    "\n",
    "        self.getFGSMTensor()\n",
    "\n",
    "    def getFGSMTensor(self):\n",
    "        self.label = K.placeholder(shape=self._model.output.get_shape())\n",
    "        self.loss = self._loss_func(self.label, self._model.output)\n",
    "        self.normed_grads = K.sign(K.gradients(self.loss, self._model.input)[0])\n",
    "        self.iterate = K.function([self._model.input, self.label], [self.loss, self.normed_grads])\n",
    "\n",
    "    def getWeights(self, model):\n",
    "        self._model.set_weights(model.get_weights())\n",
    "\n",
    "    def fgsmGrad(self, x, y, model):\n",
    "        self.getWeights(model)\n",
    "        loss_value, adv_grad = self.iterate([x, y.astype('float32')])\n",
    "        return adv_grad\n",
    "\n",
    "    def fgsm(self, x, y, model, epsilon=None):\n",
    "        adv_grad = self.fgsmGrad(x, y, model)\n",
    "        if epsilon:\n",
    "            x_adv = x+epsilon*adv_grad\n",
    "        else:\n",
    "            x_adv = x+self._epsilon*adv_grad\n",
    "        return x_adv"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "loss_func = keras.losses.sparse_categorical_crossentropy\n",
    "adv_class = AdversarialAttack(model, loss_func)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loss for prediction of the original image: 0.90025693\n",
      "Loss for prediction of the adversarial image: 1.7781217\n"
     ]
    }
   ],
   "source": [
    "epsilon = 0.1\n",
    "x_origin = x_test[8][np.newaxis, :]\n",
    "y_origin = y_test[8]\n",
    "y_pred = np.squeeze(np.argmax(model.predict(x_origin), axis=1))\n",
    "\n",
    "loss_origin = tf.Session().run(loss_func(tf.convert_to_tensor(y_origin), \n",
    "                                 tf.convert_to_tensor(model.predict(x_origin))))[0]\n",
    "\n",
    "x_adv = adv_class.fgsm(x_origin, y_origin, model, epsilon)\n",
    "y_adv = np.squeeze(np.argmax(model.predict(x_adv), axis=1))\n",
    "loss_adv = tf.Session().run(loss_func(tf.convert_to_tensor(y_origin), \n",
    "                                 tf.convert_to_tensor(model.predict(x_adv))))[0]\n",
    "\n",
    "print(\"Loss for prediction of the original image: \" + str(loss_origin))\n",
    "print(\"Loss for prediction of the adversarial image: \" + str(loss_adv))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(-0.5, 31.5, 31.5, -0.5)"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x720 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "x_origin_img = np.clip((np.squeeze(x_origin)*std+mean)*255.0, 0, 255).astype(np.int32)\n",
    "x_adv_img = np.clip((np.squeeze(x_adv)*std+mean)*255.0, 0, 255).astype(np.int32)\n",
    "plt.figure(figsize=(10, 10))\n",
    "plt.subplot(121)\n",
    "plt.imshow(x_origin_img)\n",
    "plt.axis('off')\n",
    "plt.subplot(122)\n",
    "plt.imshow(x_adv_img)\n",
    "plt.axis('off')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The truth label is: 3\n",
      "Predicted label of the original image: 3\n",
      "Predicted label of the adversarial image: 7\n"
     ]
    }
   ],
   "source": [
    "# Label from 0-9:\n",
    "# ['airplane', 'automobile', 'bird', 'cat', 'deer',\n",
    "#  'dog', 'frog', 'horse', 'ship', 'truck']\n",
    "\n",
    "print(\"The truth label is: \" + str(y_origin))\n",
    "print(\"Predicted label of the original image: \" + str(y_pred))\n",
    "print(\"Predicted label of the adversarial image: \"+ str(y_adv))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_adv_all = []\n",
    "for i in range(len(y_test)):\n",
    "    x_origin = x_test[i][np.newaxis, :]\n",
    "    y_origin = y_test[i]\n",
    "    x_adv = adv_class.fgsm(x_origin, y_origin, model, epsilon)\n",
    "    y_adv = np.squeeze(np.argmax(model.predict(x_adv), axis=1))\n",
    "    y_adv_all.append(y_adv)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test accuracy for original images: 0.5076\n",
      "Test accuracy for adversarial images: 0.1166\n"
     ]
    }
   ],
   "source": [
    "y_test = np.squeeze(y_test)\n",
    "acc_adv = sum(y_adv_all == y_test)/len(y_test)\n",
    "print(\"Test accuracy for original images: \" + str(acc))\n",
    "print(\"Test accuracy for adversarial images: \" + str(acc_adv))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}