03_Attack_GTSRB.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "mY_xKZiEnud9"
   },
   "source": [
    "### **Adversarial Attacks on GTSRB Dataset**\n",
    "\n",
    "Parameters for the **inconspicuous** attack on GTSRB data:\n",
    "* 1000 mask training epochs\n",
    "* mask threshold 0.1\n",
    "* mask regularization for noise (beta2) 0.01\n",
    "* mask deviation penalty (beta3) 4.0\n",
    "* attack LR 0.01\n",
    "* attack penalty for using colors (beta1) 1.0\n",
    "* attack deviation penalty (beta3) 4.0\n",
    "* 20000 attack training epochs\n",
    "\n",
    "\n",
    "Parameters for the **visible** attack on GTSRB data:\n",
    "* 100 mask training epochs\n",
    "* mask threshold 0.05\n",
    "* mask regularization for noise (beta2) 0.001\n",
    "* mask deviation penalty (beta3) 1.0\n",
    "* attack LR 0.0001\n",
    "* attack penalty for using colors (beta1) 0.0\n",
    "* attack deviation penalty (beta3) 1.0\n",
    "* 50000 attack training epochs\n",
    "\n",
    "\n",
    "For both attacks:\n",
    "* mask initialization LR 0.01\n",
    "* mask penalty for using colors (beta1) 0.1\n",
    "* attack batch size 8\n",
    "* attack regularization for noise (beta2) 0.0\n",
    "* augmentation is not used for initialization, but used for attack\n",
    "\n",
    "For the attack on class 14 (stop), the target class is 8 (speed limit 120km/h).\n",
    "For the attack on class 1 (speed limit 30km/h), the target class is 14 (stop).\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "sWtcKHua6jKm"
   },
   "source": [
    "## **Inconspicuous Attack**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "import os\n",
    "import json\n",
    "from models import *\n",
    "from utils import *\n",
    "\n",
    "# Parameters of LISA dataset (you can leave this as it is)\n",
    "with open('classes.json', 'r') as config:\n",
    "    params = json.load(config)\n",
    "    params['CLASS_N'] = params['GTSRB']['class_n']\n",
    "    params['DEVICE'] = params['device']\n",
    "    params['LABELS'] = params['GTSRB']['labels']\n",
    "\n",
    "    \n",
    "model_types = ['CNNsmall', 'CNNlarge', 'Transformer', 'ResNet18', 'MobileNetv2', 'EfficientNet', 'DenseNet121', 'ShuffleNetv2x1']\n",
    "\n",
    "\n",
    "model_type = \"CNNsmall\"\n",
    "\n",
    "params['MODEL_TYPE'] = model_type\n",
    "# Directory where output files will be stored\n",
    "params['OUTPUT_DIR'] = f\"attacks/inconspicuous{params['MODEL_TYPE']}GTSRB_ATTACK\"\n",
    "\n",
    "# --- PATHS INPUT---\n",
    "# Path to weights of baseline model to attack\n",
    "params['PATH_MODEL'] = f\"models/{model_type}GTSRB.pth\"\n",
    "\n",
    "# Set your hyperparameters (we add them to the params dictionary)\n",
    "\n",
    "# --- GENERAL PARAMETERS ---\n",
    "# Target class that should be recognized instead of orginal class of sign\n",
    "params['TARGET_CLASS'] = 14\n",
    "\n",
    "# Path of original sign to pertubate (should be in original printable size)\n",
    "params['PATH_SIGN'] = 'attack_utils/30kmh.jpg'\n",
    "# Path to the sign mask (defines area of sign)\n",
    "params['PATH_SIGN_MASK'] = 'attack_utils/30kmh_mask.png'\n",
    "\n",
    "# --- PATHS OUTPUT --- (ONLY CHANGE FILENAMES)\n",
    "# Path where to store the pertubated sign (printing size)\n",
    "params['PATH_PERT_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'pert_sign_small.png')\n",
    "# Path where to store the pertubated sign (printing size)\n",
    "params['PATH_PERT_SIGN_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'pert_sign_large.png')\n",
    "# Path where to store the sign in the size of the model input\n",
    "params['PATH_ORIG_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'orig_sign_small_.png')\n",
    "# Path where to store the last augmented sign\n",
    "params['PATH_AUGMENTED_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'augmented_sign_small.png')\n",
    "# Path where to store the noise in small size\n",
    "params['PATH_NOISE_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'noise_small.png')\n",
    "# Path where to store the noise in printable size\n",
    "params['PATH_NOISE_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'noise_large.png')\n",
    "# Path where to store the noise mask\n",
    "params['PATH_NOISE_MASK_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'noise_mask_small.png')\n",
    "# Path where to store the noise mask\n",
    "params['PATH_NOISE_MASK_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'noise_mask_large.png')\n",
    "# Paht where to store noise tensor\n",
    "params['PATH_NOISE_TENSOR'] = os.path.join(params['OUTPUT_DIR'], 'noise_tensor.pt')\n",
    "\n",
    "# --- PARAMETERS FOR INITIALIZATION ---\n",
    "params['INIT_MASK_THRESHOLD'] = 0.1\n",
    "# How much you want to penalty using colors\n",
    "params['INIT_MASK_BETA_1'] = 0.1\n",
    "# How much you want to regularize the ammount of noise\n",
    "params['INIT_MASK_BETA_2'] = 0.01\n",
    "# How much you want to penalty the deviation from to target class\n",
    "params['INIT_MASK_BETA_3'] = 4\n",
    "# Learning rate of adam for the mask initialization\n",
    "params['INIT_MASK_LEARNING_RATE'] = 0.01\n",
    "# How many epochs you want to train the initialization of the mask\n",
    "params['INIT_MASK_EPOCHS'] = 1000\n",
    "# Turn augmentation for initialization of mask on/off\n",
    "params['INIT_MASK_AUGMENTATION'] = False\n",
    "\n",
    "\n",
    "# PARAMETERS FOR ATTACK\n",
    "# Learning rate of adam for the attack\n",
    "params['ATTACK_LEARNING_RATE'] = 0.01\n",
    "# Batch size for one optimization step\n",
    "params['ATTACK_BATCH_SIZE'] = 8\n",
    "# How much you want to penalty using colors in the attack\n",
    "params['ATTACK_BETA_1'] = 1\n",
    "# How much you want to regularize the ammount of noise in the attack\n",
    "params['ATTACK_BETA_2'] = 0.0\n",
    "# How much you want to penalty the deviation from to target class\n",
    "params['ATTACK_BETA_3'] = 4\n",
    "# How many epochs you want to train the attack\n",
    "params['ATTACK_EPOCHS'] = 20000\n",
    "# Turn augmentation on/off for attack\n",
    "params['ATTACK_AUGMENTATION'] = True\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(model_type)\n",
    "transform_to_tensor = transforms.Compose([transforms.Resize((32, 32)),\n",
    "                                              transforms.ToTensor()])\n",
    "\n",
    "orig_sign = torch.clamp(transform_to_tensor(Image.open(params['PATH_SIGN'])), 0, 1)\n",
    "\n",
    "model = load_model(params)\n",
    "\n",
    "print(\"Prediction on the original sign:\")\n",
    "test_on_tensor(model=model, tensor=orig_sign, params=params, labels=params['LABELS'])\n",
    "print(\"Prediction for the stop sign:\")\n",
    "test_on_tensor(model=model, tensor=orig_sign, params=params, labels=params['LABELS'], specific_class=14)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 1000,
     "output_embedded_package_id": "149RTv_3jGN4UCT8yBXbx1l2twrUA2trj"
    },
    "executionInfo": {
     "elapsed": 4027338,
     "status": "ok",
     "timestamp": 1689198485039,
     "user": {
      "displayName": "Leopold Müller",
      "userId": "15867647540029964791"
     },
     "user_tz": -120
    },
    "id": "hQbOfUfI6sTp",
    "outputId": "d49ff19a-3747-4621-e7df-a352a115052e"
   },
   "outputs": [],
   "source": [
    "# Create output directory\n",
    "if not os.path.exists(params['OUTPUT_DIR']):\n",
    "  os.makedirs(params['OUTPUT_DIR'])\n",
    "\n",
    "print(f\"-----------------{model_type}----------------------\")\n",
    "\n",
    "# Train attack\n",
    "train_attack(params)\n",
    "# Test attack\n",
    "test_attack(params)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "8rVJsSpT-R56"
   },
   "source": [
    "## **Visible Attack**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 1000,
     "output_embedded_package_id": "1cXo16mjn7ZFWxRpyfXKD59XETr-OfFux"
    },
    "executionInfo": {
     "elapsed": 3947446,
     "status": "ok",
     "timestamp": 1689097614106,
     "user": {
      "displayName": "Leopold Müller",
      "userId": "15867647540029964791"
     },
     "user_tz": -120
    },
    "id": "QsC4pOsj9uK1",
    "outputId": "b5ad4034-6634-4564-ab8f-e4fe894f1dcf"
   },
   "outputs": [],
   "source": [
    "import os\n",
    "import json\n",
    "from models import *\n",
    "from utils import *\n",
    "\n",
    "# Parameters of LISA dataset (you can leave this as it is)\n",
    "with open('classes.json', 'r') as config:\n",
    "    params = json.load(config)\n",
    "    params['CLASS_N'] = params['GTSRB']['class_n']\n",
    "    params['DEVICE'] = params['device']\n",
    "    params['LABELS'] = params['GTSRB']['labels']\n",
    "\n",
    "model_types = ['CNNsmall', 'CNNlarge', 'Transformer', 'ResNet18', 'MobileNetv2', 'EfficientNet', 'DenseNet121', 'ShuffleNetv2x1']\n",
    "\n",
    "model_type = \"DenseNet121\"\n",
    "\n",
    "\n",
    "# Set your hyperparameters (we add them to the params dictionary)\n",
    "\n",
    "# --- GENERAL PARAMETERS ---\n",
    "# Target class that should be recognized instead of orginal class of sign\n",
    "params['TARGET_CLASS'] = 14\n",
    "params['MODEL_TYPE'] = model_type\n",
    "# Directory where output files will be stored\n",
    "params['OUTPUT_DIR'] = f\"attacks/visible{params['MODEL_TYPE']}GTSRB_ATTACK\"\n",
    "\n",
    "# --- PATHS INPUT---\n",
    "# Path to weights of baseline model to attack\n",
    "params['PATH_MODEL'] = f\"models/{model_type}GTSRB.pth\"\n",
    "# Path of original sign to pertubate (should be in original printable size)\n",
    "params['PATH_SIGN'] = 'attack_utils/30kmh.jpg'\n",
    "# Path to the sign mask (defines area of sign)\n",
    "params['PATH_SIGN_MASK'] = 'attack_utils/30kmh_mask.png'\n",
    "\n",
    "\n",
    "# --- PATHS OUTPUT --- (ONLY CHANGE FILENAMES)\n",
    "# Path where to store the pertubated sign (printing size)\n",
    "params['PATH_PERT_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'pert_sign_small.png')\n",
    "# Path where to store the pertubated sign (printing size)\n",
    "params['PATH_PERT_SIGN_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'pert_sign_large.png')\n",
    "# Path where to store the sign in the size of the model input\n",
    "params['PATH_ORIG_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'orig_sign_small_.png')\n",
    "# Path where to store the last augmented sign\n",
    "params['PATH_AUGMENTED_SIGN_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'augmented_sign_small.png')\n",
    "# Path where to store the noise in small size\n",
    "params['PATH_NOISE_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'noise_small.png')\n",
    "# Path where to store the noise in printable size\n",
    "params['PATH_NOISE_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'noise_large.png')\n",
    "# Path where to store the noise mask\n",
    "params['PATH_NOISE_MASK_SMALL'] = os.path.join(params['OUTPUT_DIR'], 'noise_mask_small.png')\n",
    "# Path where to store the noise mask\n",
    "params['PATH_NOISE_MASK_LARGE'] = os.path.join(params['OUTPUT_DIR'], 'noise_mask_large.png')\n",
    "# Paht where to store noise tensor\n",
    "params['PATH_NOISE_TENSOR'] = os.path.join(params['OUTPUT_DIR'], 'noise_tensor.pt')\n",
    "\n",
    "# --- PARAMETERS FOR INITIALIZATION ---\n",
    "params['INIT_MASK_THRESHOLD'] = 0.01\n",
    "# How much you want to penalty using colors\n",
    "params['INIT_MASK_BETA_1'] = 0\n",
    "# How much you want to regularize the ammount of noise\n",
    "params['INIT_MASK_BETA_2'] = 0\n",
    "# How much you want to penalty the deviation from to target class\n",
    "params['INIT_MASK_BETA_3'] = 1\n",
    "# Learning rate of adam for the mask initialization\n",
    "params['INIT_MASK_LEARNING_RATE'] = 0.01\n",
    "# How many epochs you want to train the initialization of the mask\n",
    "params['INIT_MASK_EPOCHS'] = 1000\n",
    "# Turn augmentation for initialization of mask on/off\n",
    "params['INIT_MASK_AUGMENTATION'] = False\n",
    "\n",
    "\n",
    "# PARAMETERS FOR ATTACK\n",
    "# Learning rate of adam for the attack\n",
    "params['ATTACK_LEARNING_RATE'] = 0.0001\n",
    "# Batch size for one optimization step\n",
    "params['ATTACK_BATCH_SIZE'] = 8\n",
    "# How much you want to penalty using colors in the attack\n",
    "params['ATTACK_BETA_1'] = 0.0001\n",
    "# How much you want to regularize the ammount of noise in the attack\n",
    "params['ATTACK_BETA_2'] = 0.0\n",
    "# How much you want to penalty the deviation from to target class\n",
    "params['ATTACK_BETA_3'] = 1.0\n",
    "# How many epochs you want to train the attack\n",
    "params['ATTACK_EPOCHS'] = 20000\n",
    "# Turn augmentation on/off for attack\n",
    "params['ATTACK_AUGMENTATION'] = True\n",
    "\n",
    "# Create output directory\n",
    "if not os.path.exists(params['OUTPUT_DIR']):\n",
    "  os.makedirs(params['OUTPUT_DIR'])\n",
    "\n",
    "print(f\"-----------------{model_type}----------------------\")\n",
    "\n",
    "# Train attack\n",
    "train_attack(params)\n",
    "# Test attack\n",
    "test_attack(params)"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [
    "d8d_BzabIAXt",
    "q2khGChjH6dW",
    "QgiIk-8LnueC"
   ],
   "gpuType": "T4",
   "provenance": []
  },
  "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": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}