00-sanity-checks.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import sys\n",
    "sys.path.insert(0, \"./src/\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 00-Sanity-checks"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Run sanity checks for attacks and data modification."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "c:\\Users\\denkm\\anaconda3\\envs\\practical_work\\Lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
      "  from .autonotebook import tqdm as notebook_tqdm\n"
     ]
    }
   ],
   "source": [
    "import tqdm\n",
    "import os\n",
    "\n",
    "import numpy as np\n",
    "import torch\n",
    "import torchvision\n",
    "import torch.nn as nn\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from src.utils.data_utils import get_cifar10_data\n",
    "from src.utils.adversary_utils import adv_attack\n",
    "from src.utils.vis_utils import vis_pred\n",
    "\n",
    "torch.manual_seed(0)\n",
    "torch.backends.cudnn.deterministic = True"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Files already downloaded and verified\n",
      "Files already downloaded and verified\n"
     ]
    }
   ],
   "source": [
    "data = get_cifar10_data(32, img_size=32, subsample=True)\n",
    "classes = data[\"classes\"]\n",
    "_, val_laoder = data[\"val\"]\n",
    "_, test_loader = data[\"test\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = torchvision.models.resnet50()\n",
    "model.fc = nn.Linear(2048, len(classes))\n",
    "model.load_state_dict(torch.load(os.path.join(r\"ckpts\\standard_d_v15\\best.pt\")))\n",
    "#model.fc = nn.Identity()\n",
    "model.to(\"cuda\");\n",
    "model.eval();"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "adversary_param_linf = {\"steps\": 1, #1000,\n",
    "                           \"epsilon\": 8/255,\n",
    "                           \"alpha\": 2/255}\n",
    "\n",
    "adversary_param_l2 = {\"steps\": 10,\n",
    "                      \"epsilon\": 0.5,\n",
    "                      \"alpha\": 0.1}\n",
    "\n",
    "adversary_param = adversary_param_linf"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Closer look at attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Files already downloaded and verified\n",
      "Files already downloaded and verified\n"
     ]
    }
   ],
   "source": [
    "data = get_cifar10_data(1, img_size=32, subsample=True)\n",
    "classes = data[\"classes\"]\n",
    "_, val_laoder = data[\"val\"]\n",
    "_, test_loader = data[\"test\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = torchvision.models.resnet50()\n",
    "model.fc = nn.Linear(2048, len(classes))\n",
    "model.load_state_dict(torch.load(os.path.join(r\"ckpts\\standard_d_v15\\best.pt\")))\n",
    "model.to(\"cuda\");\n",
    "model.eval();"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True label:  cat\n",
      "Prediced label on original image:  cat\n",
      "Prediction after adv attack 1:  dog\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True label:  ship\n",
      "Prediced label on original image:  ship\n",
      "Prediction after adv attack 1:  car\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True label:  ship\n",
      "Prediced label on original image:  car\n",
      "Prediction after adv attack 1:  car\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True label:  plane\n",
      "Prediced label on original image:  plane\n",
      "Prediction after adv attack 1:  cat\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAaEAAAGdCAYAAAC7EMwUAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8pXeV/AAAACXBIWXMAAA9hAAAPYQGoP6dpAAAtrElEQVR4nO3dfXCVdX738c91nac8cBJAzJNgzCroCkq3YhXqKtKSMZ06umzndteZHbzbOusqzjDsji36h5nOlHjbkXFnqLTd7lidavGPqnVGV2VHgd2bZQsWC4vWxRXXrBBZEZKQh/N0/e4/rLkbAf19IfGXhPdr5szIOV+/+V0P53xzJed8EjnnnAAACCAOvQAAwNmLIQQACIYhBAAIhiEEAAiGIQQACIYhBAAIhiEEAAiGIQQACCYdegGfliSJDh48qHw+ryiKQi8HAGDknFN/f79aWloUx599rTPhhtDBgwc1Z86c0MsAAJyh7u5uzZ49+zNrxm0IPfLII/qbv/kbHTp0SPPnz9fDDz+sr371q5/7/+XzeUnS//k/96m6usrra9XO+JL3uuqqp3nXSlJtfrp37bHjZVPvweNHvWuzKVu6UiWpeNdGke2nstXZalN9LpXxLzZuZ+IS/2LjhXVS8d+HiTH9KqnYzhVL/1zGsL8lRUr5r0OG/S0pnfJ/iYmNvZ0s57itt/V4RjIcnyq/17VPpKOcd20c15h6xyn/J8XxD9/wrh0cHNL/+t+rRl7PP8u4DKGnnnpKq1ev1iOPPKLf//3f19///d+ro6NDb7zxhs4///zP/H8/+RFcdXWV9xCqqfHf8TXVtd61klRb6z+0iq5k6q2k6F2aTdmeROM7hGwnelV6cg6himUIJeM9hPy3syqbNfU2DSHL/paUTjOEPm08h1Aqtr2+xWn/J4UbtD3vJXn9SmVc3piwfv16/dmf/Zn+/M//XF/+8pf18MMPa86cOdq4ceN4fDkAwCQ15kOoWCzqtddeU3t7+6j729vbtX379hPqC4WC+vr6Rt0AAGeHMR9CH374oSqVihobG0fd39jYqJ6enhPqu7q6VF9fP3LjTQkAcPYYt88Jffpngc65k/58cO3atert7R25dXd3j9eSAAATzJi/MWHWrFlKpVInXPUcPnz4hKsjScrlcsrl/H/xBgCYOsb8SiibzeqKK67Q5s2bR92/efNmLVmyZKy/HABgEhuXt2ivWbNG3/rWt7Ro0SItXrxY//AP/6D33ntPd9xxx3h8OQDAJDUuQ+iWW27RkSNH9Fd/9Vc6dOiQFixYoBdeeEGtra3j8eUAAJPUuCUm3HnnnbrzzjtPv0Gck2K/D3W5qlnebUvZetMyKqnp3rWZjP+HTyWpMHTcv7YyYOqdJIYPzkb+H1aUpFLO9sG/QUP7dJXtE6WFkv8+T2dsp/vQ8UHvWusHfrPGVINSyX+fF1K2D8JWKv7nijN+oLR2mv+HJwsF24e9yyX/D6vmjB/gjWPbc2LGDP/XlZraGabeLjJ8cNbw4VNJSmX9nxOVXv9jWSn7r4MUbQBAMAwhAEAwDCEAQDAMIQBAMAwhAEAwDCEAQDAMIQBAMAwhAEAwDCEAQDAMIQBAMOMW23OmMpm0Mp4xK+XYf5YOJP5RH5KUdcPetdXTbbtzRubEP21xKtFHH5l6V/f1etdWKoZYEEmV2mpTfaF2mndtPmeLHZkh/797n04Z44ny/ut2tl2omhrbPnSJ/xdIElu0jmTY55Gtd1WV/3YODPjHJElSsegfT1RVZYvtSRlje6qr/CLGPu5taq1yYogDM5wnkhTH/s8fGV5nLbVcCQEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCmbDZcUPlWqnslzuVrvjnQmWNW1yUf+84ss306lTGu7auxrbwaM8e79q+j/zz8STpnIsvNNUr9t/OUuyf1yZJdSn/3LPjg/2m3mlDDld15J8dJkmpWXlTfTJU8K6NY1u+WyHvnx+WKtqyF3PKedcOpG3HZ9qxo961+blfMfU+PmO6qT4pDHnXFp1/5p0kZcr+z/3Iel1R9n9uxhX/PL24QnYcAGASYAgBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCmbCxPUlJSjxXZ4pXiW2xI67kH4FSjPwjZCQpSvwjgYZT9abemXKtd21leqOp92B/0VRffvct71qX8o+QkaTIPxVGx1O2dceJ/7HPlmyxPcPv+UegSFJS8j9vo8j/+SBJhTrDPh8smXpXO0Nk0zmm1ir0fORdm60+39Y8P8tUXra8TmRsr0FpQ7yXky0SKI7869Mp//MqbTi9uRICAATDEAIABMMQAgAEwxACAATDEAIABMMQAgAEwxACAATDEAIABMMQAgAEwxACAATDEAIABDNhs+MqFansGbGUJIbMtsi2ya7in5c0nNhytTJZQzbZgC1vanj6TO/aeJYtV6ucHjTVx1X+QVLlc5tNvQfT/vsw+s0Hpt5xxj+zq7+62tQ7abRlk2Wc/z4cTmz5YdV5/5zBoWPHTb0rVf7hfqmM7XviqH/AuzZusuUj5rK251sl9s8OnJ6yZUym5H+Ol63XFbH/OR6Zav1fZ7kSAgAEM+ZDqLOzU1EUjbo1NTWN9ZcBAEwB4/LjuPnz5+vHP/7xyL9TKVtsPQDg7DAuQyidTnP1AwD4XOPyO6H9+/erpaVFbW1t+sY3vqF33nnnlLWFQkF9fX2jbgCAs8OYD6GrrrpKjz/+uF566SX94Ac/UE9Pj5YsWaIjR46ctL6rq0v19fUjtzlz5oz1kgAAE9SYD6GOjg59/etf12WXXaY//MM/1PPPPy9Jeuyxx05av3btWvX29o7curu7x3pJAIAJatw/J1RbW6vLLrtM+/fvP+njuVxOuZz/ZwkAAFPHuH9OqFAo6M0331Rzs+1DiACAqW/Mh9D3vvc9bd26VQcOHNDPf/5z/cmf/In6+vq0cuXKsf5SAIBJbsx/HPeb3/xG3/zmN/Xhhx/q3HPP1dVXX60dO3aotbXV1CdxiRLnF1dRKA57980VbZEmqbR/HEtStsVxOOe/luwpfpx5KpW9O/zXsci2T+KULaKmUqnxrk0ds0UfDUf+0S1V7x819c7V5b1rKzX+8U6SlJYt/saVi96106bXmXrn+j/yro37bOtWwzTv0ugDWxxUqs4/bmhw3+um3pm8/7GXpGTefO/a4Thr6h0ZYnvSJdtrULro3zsxJBlZasd8CG3atGmsWwIApiiy4wAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwYz7n3I4Xak4rXSc8aqNXcq7b6Xsn5UkSRXnH4KUSvut9xO1R/xzuErdB029Z2T8892O93xo6l2sqjfVl4qG73UGPzD1rpntnx82XG/Ld3NZ/0zCbN+QqXf2aL+pvhD5Z+oVDplaq6rsv1+S4/5ZfZKU+q1/bmBS8H8eS1JV3Tzv2qO/tj1/lLblu9XNOt+7NlPrn6cnSZXI/zVoWNYMQ//XrKLz722p5UoIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABDMhI3tyU/Lq7raL/IjqvKPqElFtniVSJF3bWyIqpCkwSr/mJLBRQtNvYdTV/j3Pm6LkCkZz5qKoT4yxipVT/OPhekv2Y59KvZfeLZgaq2qjC0WZjDrfx5GFf+IH0kqRf7fix7v94+akmzbWcravidOGmZ5186aZoua6i/Z9mGfYe2pjO11oibxPw8T43VFOfZfiyHdyVTLlRAAIBiGEAAgGIYQACAYhhAAIBiGEAAgGIYQACAYhhAAIBiGEAAgGIYQACAYhhAAIBiGEAAgmAmbHTettlo1NX65YKWc/2YUS2XTOqK0f30lseWeJZmcd21NY52pd99x/5y0nmN9pt6ZrC33rFAc9q7Nxf55epL0Yd+Ad+2Q8djnMv55bVmXMfWujmzbWfZfilSxFEsflvyD75LE9n1rplTxri0WbHltgzn/82re7Dmm3jXGc6UU+W+nc7bjI8NzwnxVYci7TGTYRkMtV0IAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYCZsdlyccopTfrlG1TVV3n2TIdsmp7P+mWBJZMuOyxjylaKKfxacJFWcf30q678OSUqnbNuZjv2/1ykO95t6Z9P+x6e6ypZ5lzH0Tke27LhKuWiq17Ch3nh8qmr8114p2XrXZPyPfdk/wk6SlAz475N+Yy5dVdb2OlFryWAr23Lp5Pk6KEkV2+FRnPI/PpEh885Sy5UQACAY8xDatm2bbrzxRrW0tCiKIj377LOjHnfOqbOzUy0tLaqurtbSpUu1b9++sVovAGAKMQ+hgYEBLVy4UBs2bDjp4w8++KDWr1+vDRs2aOfOnWpqatLy5cvV32/7MQsAYOoz/06oo6NDHR0dJ33MOaeHH35Y9913n1asWCFJeuyxx9TY2Kgnn3xS3/72t89stQCAKWVMfyd04MAB9fT0qL29feS+XC6n6667Ttu3bz/p/1MoFNTX1zfqBgA4O4zpEOrp6ZEkNTY2jrq/sbFx5LFP6+rqUn19/chtzhzbX0AEAExe4/LuuCga/fY859wJ931i7dq16u3tHbl1d3ePx5IAABPQmH5OqKmpSdLHV0TNzc0j9x8+fPiEq6NP5HI55XK5sVwGAGCSGNMroba2NjU1NWnz5s0j9xWLRW3dulVLliwZyy8FAJgCzFdCx48f19tvvz3y7wMHDuj111/XzJkzdf7552v16tVat26d5s6dq7lz52rdunWqqanRrbfeOqYLBwBMfuYhtGvXLl1//fUj/16zZo0kaeXKlfqnf/on3XPPPRoaGtKdd96po0eP6qqrrtLLL7+sfD5v+jrZTFpZz+iMTI1/HEucqjatoyZT411bcrbckf5e//pUlDL1rpox07u2IV9n6i35x4hIkkv8Y0rSxmvzdGw49pHtdM8YelcS2z5Jxbb6dGSIhZEtFqZoiJtKEmPEk+FcySa2gz847B/FM5DYnptlYwRX1lBfjo2xPYYfWDlne51I+6frKJ32750y1JqH0NKlS+U+IycpiiJ1dnaqs7PT2hoAcJYhOw4AEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEMyY/imHsZSKK0rHfjlVaRX9Gxvzwz7q+dC79sjxg6beRw69711bn59h6n3Z/Mu8a9NVtjy9ocSWq1Uo+x+fTGz7vigdZfyLjb0zKf/8q6jkn2MmSbEhC06SIkPWnKvYeicF/zy4KGUIG5MUpf2Pj8vYeqdP8TfKTiZlzDtMOdu5EhnOrWzKP5NQkjKG7TQuW+mU/+uhM6yjYqjlSggAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEMyEje2Ri70zKOLIf5Ya01LU19vnXfvbIz2m3seO+cf2/Nfefzf1fusXO7xrL/rSpabebRfON9VPP6fBvzj2j8qRpKFCwbs2nbb1zhoiajJpW+SMnDW2x79/OmWIMpKUzfjXl42RTYlhOzM528tROmuIBDLu70rZP8pIkpQYYpWcbR+WTOW28zCJ/LdzqOAfTTVU9K/lSggAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQDEMIABAMQwgAEAxDCAAQzITNjosip8gz6C2d9c8ES0VVpnVc+uVLvGsvis8z9R7o/8C7dt/u3abe//kfP/eu/cmr75l6/9d//sJU/+X5X/GunbfgMlPvXI3/8cxmbKd7dc7/vEqMWWORbDl2xbJ/gFjKGJBYNOSqJeWyqbcls600aOudOP+ctES2vDZr1pwzZLZFxpjB2LAWZzyvYkNWY8Vynhh2H1dCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgJmxsT5I4JZ7ZD3HsvxkuZYvviA1xHKlU1tS7fuZs79qvXt9k6v3leRd5127d/GNT77ffetdUPzA45F07WDhu6j3v0ku9a2ef32rqnST+MTKFgv82SlI6nTPVVyolQ60tciaO/KNbYmPkjGJD3JDhuSZJqdj/e+gksj03rRJTpI014slfJbFdVySJf/eKISbJcgpyJQQACIYhBAAIxjyEtm3bphtvvFEtLS2KokjPPvvsqMdvu+02RVE06nb11VeP1XoBAFOIeQgNDAxo4cKF2rBhwylrbrjhBh06dGjk9sILL5zRIgEAU5P5jQkdHR3q6Oj4zJpcLqemJtsv0gEAZ59x+Z3Qli1b1NDQoHnz5un222/X4cOHT1lbKBTU19c36gYAODuM+RDq6OjQE088oVdeeUUPPfSQdu7cqWXLlqlQKJy0vqurS/X19SO3OXPmjPWSAAAT1Jh/TuiWW24Z+e8FCxZo0aJFam1t1fPPP68VK1acUL927VqtWbNm5N99fX0MIgA4S4z7h1Wbm5vV2tqq/fv3n/TxXC6nXM72wT0AwNQw7p8TOnLkiLq7u9Xc3DzeXwoAMMmYr4SOHz+ut99+e+TfBw4c0Ouvv66ZM2dq5syZ6uzs1Ne//nU1Nzfr3Xff1b333qtZs2bpa1/72pguHAAw+ZmH0K5du3T99deP/PuT3+esXLlSGzdu1N69e/X444/r2LFjam5u1vXXX6+nnnpK+Xze9HWiVEpRyi/TypIdF2dO/gaJU8kZ9lDasA5JSir+2WTptK33hXMv964tDppa6/3uTab6I7/9jXftf+21Lebg+7/0rr1g3pdNvS/60lzv2vr66abe07LWp56h3tnyEUueGY2SpMiW7xbF/rl05bItU03yz9MzxtIpyhiz5iz73LiWOPL/gVVk3IfOcjwjwzloqDUPoaVLl8p9RljfSy+9ZG0JADhLkR0HAAiGIQQACIYhBAAIhiEEAAiGIQQACIYhBAAIhiEEAAiGIQQACIYhBAAIhiEEAAhm3P+Uw+mK5B+xFBvyj9KRISdLUtYQrRQbcrIkKZF/zlNkDJwaGi561zafN9vU+8KLvmSq3/mz971rS2XbX9btOXjUu/Y3Pe+Zer+x5z+8ay+66EJT77YLbPvw3HP8U+jz+XpT77Qy3rUF/9NKkmR5tuWqqk29CyX/HMhiYsvTS8W2+ij231JnXEsSGV4nnO11wvTaaXgNShlquRICAATDEAIABMMQAgAEwxACAATDEAIABMMQAgAEwxACAATDEAIABMMQAgAEwxACAAQzYWN74qii2DOuIjbE3yTlkmkdlciQUxLbIjNc4r/uVMo/WkWSIpW9a6trqky9p58zw1SfrvI/zcoV/30iSUPDw961LrId+wH91rv2F7s/MvV+0xAJJEnT6/PetS2NLabe58w637s2ztoigc5t8I+EOv8C/3VIUnWN/3NiwHCeSFLZGO+VGJ7LZeNrkEUkW3RYZIgQchVD3JChlishAEAwDCEAQDAMIQBAMAwhAEAwDCEAQDAMIQBAMAwhAEAwDCEAQDAMIQBAMAwhAEAwDCEAQDATODvOKfbMb8qm/GdpktgyoWLnn4GUOFtvGXrL2DtXlfOuHeg9Zurd03PQVH/w0CHv2uqc7ZRMpfy3s66uztQ7LluysmzHZ7gyZKo/ONzrXfvr935l6t1/3D8fcaBkyyarm9HkXXvlooWm3pfOv8C7dub0c0y9q/O2+lyVf6ZeNlVt6m3Jd0sSW35lFPm/dg4bTnFL9B5XQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYCZsbE+ksiKVvGoTS6xFOWtaR0X+vSsyxvbE/mupVIytDZFAe/7jNVPvY7/9wFRfX1vrXXvogyOm3nX1Ge/aGv90GklSX2nAuzaTMcYNZWzxN6lMjXdtOq4y9S4n/d61g32Dpt4HD77tXfvqlvdNvV97zf/5U52x7ZPzZn/JVN/SfIF/7exWW++G871ra2qnm3qr2v88jAwRWVGq7F3LlRAAIBjTEOrq6tKVV16pfD6vhoYG3XzzzXrrrbdG1Tjn1NnZqZaWFlVXV2vp0qXat2/fmC4aADA1mIbQ1q1bddddd2nHjh3avHmzyuWy2tvbNTDw/39s8eCDD2r9+vXasGGDdu7cqaamJi1fvlz9/f6X/ACAs4PpB9kvvvjiqH8/+uijamho0GuvvaZrr71Wzjk9/PDDuu+++7RixQpJ0mOPPabGxkY9+eST+va3vz12KwcATHpn9Duh3t6P/8bJzJkzJUkHDhxQT0+P2tvbR2pyuZyuu+46bd++/aQ9CoWC+vr6Rt0AAGeH0x5CzjmtWbNG11xzjRYsWCBJ6unpkSQ1NjaOqm1sbBx57NO6urpUX18/cpszZ87pLgkAMMmc9hBatWqV9uzZo3/5l3854bEoGv3X/ZxzJ9z3ibVr16q3t3fk1t3dfbpLAgBMMqf1OaG7775bzz33nLZt26bZs2eP3N/U9PGf8u3p6VFzc/PI/YcPHz7h6ugTuVxOuZz/+88BAFOH6UrIOadVq1bp6aef1iuvvKK2trZRj7e1tampqUmbN28eua9YLGrr1q1asmTJ2KwYADBlmK6E7rrrLj355JP6t3/7N+Xz+ZHf89TX16u6ulpRFGn16tVat26d5s6dq7lz52rdunWqqanRrbfeOi4bAACYvExDaOPGjZKkpUuXjrr/0Ucf1W233SZJuueeezQ0NKQ777xTR48e1VVXXaWXX35Z+Xx+TBYMAJg6TEPIuc/PRouiSJ2dners7DzdNUmSiuWCMmW/XKNCwRAKVjJmfKX8M9hcZMuOc84/EC59ijd2nEp/r/9b3YcHhk29L553qan+K5df4V27e+8bpt47du3yrv3tR8dMvZNKwbv28t/5HVPva6+91lSfrfHPPjvwy1+aev/fHVu9ay+5ZJ6p98zG5s8v+m+/+fWvTb0PHTzoXVtO+R9LSWo2rFuS2i7wz4Mr+7+kSJL6e49511YMOZqSlEn7ZxIOl/xfrwqGWrLjAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBnNafcvgiuOTjm5fIMEtt6Tem+lTKFttTcWXv2sQjMul/ylX5x7xcu2y5qXccGU+bxH/tF3/lalPrBYuuNKzDP0pEktJp/+2sq5tp6n1B64Wm+ny9f/biBW0LTL0bW/2jeDJZ259dOWfWLO/aivH4HDnyW/9i29NHDbNO/qdnTiVfW+9dG6Vsz5+0/OLLJKmS2OKJSoY4sCT234mWWq6EAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMFM2Oy4wcFB79q4t9+7NmPIYZKkQlL0rq1EtuyrStm/vljwX4ckJRX/3ql01tS7XPHPvJOkSuIbAii52Bbud94F/hlsiWF/S1I67Z+TVhq2hZP96t0jpvpy5SPv2kzO9rSeMcs/O27YeB4ePuJfn8lVm3rXz5jrXRsbn/dHevtM9QcP93rXliv+zwdJqooz3rW5jO08jGr998vw0SHv2qGhYe9aroQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMFM2Nie7dt/qlzOL07mWHmPd99p6VrTOsrDA/61hngaSaq4kn/vcsHW2xBRU0mMcUPOtp1R7H+alRNbbE+p6L8PVbH1rqqZ5l07PT/L1Lu2us5UXzJsplK27y2rspbjY4uFMT0nUraXo+pMjXetJYJJkkplWzRVKu0freNk24dpZ6g3vKZIUlTlf67Ewx961xYKxPYAACYBhhAAIBiGEAAgGIYQACAYhhAAIBiGEAAgGIYQACAYhhAAIBiGEAAgGIYQACAYhhAAIJgJmx1Xk61WLuuX91RK+WdIpRLbJlfl6r1rk8g20y0JbCljHpgM+W5JYsvJcs6WwVa29E/bcrVkyDKzRHBJUjrtl10oSeWCLX8vHflna0lSnPU/nkOFIVPviuF70Ti2nYeRIcOwMGTbh0nKP68tilO23s62lkzGv7/1uTxwvNe7tjxky5gcNjyVq7LHvGuLxaJ3LVdCAIBgTEOoq6tLV155pfL5vBoaGnTzzTfrrbfeGlVz2223KYqiUberr756TBcNAJgaTENo69atuuuuu7Rjxw5t3rxZ5XJZ7e3tGhgY/ecObrjhBh06dGjk9sILL4zpogEAU4PpFyQvvvjiqH8/+uijamho0GuvvaZrr7125P5cLqempqaxWSEAYMo6o98J9fZ+/AuzmTNnjrp/y5Ytamho0Lx583T77bfr8OHDp+xRKBTU19c36gYAODuc9hByzmnNmjW65pprtGDBgpH7Ozo69MQTT+iVV17RQw89pJ07d2rZsmUqFE7+ro2uri7V19eP3ObMmXO6SwIATDKn/RbtVatWac+ePfrpT3866v5bbrll5L8XLFigRYsWqbW1Vc8//7xWrFhxQp+1a9dqzZo1I//u6+tjEAHAWeK0htDdd9+t5557Ttu2bdPs2bM/s7a5uVmtra3av3//SR/P5XLK5Wx//x0AMDWYhpBzTnfffbeeeeYZbdmyRW1tbZ/7/xw5ckTd3d1qbm4+7UUCAKYm0++E7rrrLv3zP/+znnzySeXzefX09Kinp0dDQx9/Qvv48eP63ve+p5/97Gd69913tWXLFt14442aNWuWvva1r43LBgAAJi/TldDGjRslSUuXLh11/6OPPqrbbrtNqVRKe/fu1eOPP65jx46publZ119/vZ566inl8/kxWzQAYGow/zjus1RXV+ull146owV9IqkUlXjGNw0cP+bdtzZl+/1TZAgcKxuzycoV/3yqYtmWNaaKf4ZUFNuy4yqJfy6UJA2XSt61pYrt15QVS3ZcxZLWJ1kS8hLDOiQpimxrcYn/8Swas+OSxH8tTrbtTEqG7TTuk8Tw3ExscYdKp2z/Qzrlv5ZS2f/5IEmRIYetXLJl3h2v8784OK9tundtKfLfRrLjAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBnPbfExpv77//pjIZv+X96gP/iIjqVMa0jnTFv3fZECMiSVGU9a6tOFukicr+UR/ZrO17kXLFFttjCSkpW7OPIv94ldgYxRIbYmTSKds+tETlSFIqTnnXGvegymX/qJeKoVayRR+lY9vLkeX5k1Rs0VSl2LYXC74ZY7Ltb0kqRf71ccMMU++m+Qs+v+i/1df6H83isP9rBFdCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAmbHackoyU+C0vY5ilkRu/jK8osvW2lMcVY2ZX2j8jzxpLFyW2/L1M2nCaGTPYLAlfzpgf5gz5YdZsv1TK9tRLYv/cLmv8XtlwIlZS/hl2khQZ1uIyxt6J//GMne35Y4yBVMmwX8o5/8w7SaqbPcu7tvV35pt6+77GStLRt/Z615aK/omRXAkBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCYQgBAIJhCAEAgmEIAQCCYQgBAIKZsLE9xWJBzjNqozRc9G9clbOto1TwLzZEYEhSKusfxZIYY0ciGeKGjJEzcWysN4TrFMu23umUfwRKZEuFGddv0Zz/oZckJWXL8TduaOK/mHRki2yKDTFMqdgYrSP/aJiq2HYwixXjwa+t9S6deck8U+vzLpjtXVv44ANT7wP7fuFdW1Me9K4tl/yPJVdCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAYQgCAYBhCAIBgGEIAgGAmbHacUvKOwIrS/tlkUca/VpKysX8Ol0tsmV2xIePLOVtvSzSZi8qm3sYYLmXS/nljM+unmXontqg5W29DXl8mZTyvMrYMtoplO53taV027MQkseW7DQwMeNdWjAfTGXIDjxrP2bhxuql+9txLvGvPmdlg6n3wzV961x59+1em3umK/3O/ynDKlgwvKVwJAQCCMQ2hjRs36vLLL1ddXZ3q6uq0ePFi/ehHPxp53Dmnzs5OtbS0qLq6WkuXLtW+ffvGfNEAgKnBNIRmz56tBx54QLt27dKuXbu0bNky3XTTTSOD5sEHH9T69eu1YcMG7dy5U01NTVq+fLn6+/vHZfEAgMnNNIRuvPFG/dEf/ZHmzZunefPm6a//+q81bdo07dixQ845Pfzww7rvvvu0YsUKLViwQI899pgGBwf15JNPjtf6AQCT2Gn/TqhSqWjTpk0aGBjQ4sWLdeDAAfX09Ki9vX2kJpfL6brrrtP27dtP2adQKKivr2/UDQBwdjAPob1792ratGnK5XK644479Mwzz+jSSy9VT0+PJKmxsXFUfWNj48hjJ9PV1aX6+vqR25w5c6xLAgBMUuYhdPHFF+v111/Xjh079J3vfEcrV67UG2+8MfJ4FI1+c7Bz7oT7/qe1a9eqt7d35Nbd3W1dEgBgkjJ/Tiibzeqiiy6SJC1atEg7d+7U97//ff3FX/yFJKmnp0fNzc0j9YcPHz7h6uh/yuVyyuVy1mUAAKaAM/6ckHNOhUJBbW1tampq0ubNm0ceKxaL2rp1q5YsWXKmXwYAMAWZroTuvfdedXR0aM6cOerv79emTZu0ZcsWvfjii4qiSKtXr9a6des0d+5czZ07V+vWrVNNTY1uvfXW8Vo/AGASMw2hDz74QN/61rd06NAh1dfX6/LLL9eLL76o5cuXS5LuueceDQ0N6c4779TRo0d11VVX6eWXX1Y+nzcvLJOklUn8lhcVh737JlHBtA6XFL1rU7JFsZjqP+P3aifj5B+vEskWlxJHtvpK0X8fHjdGt0Sx/ykcp6z70D+KZ7hUMvVOFY0xTJGh3tl+wPFZv7M9odbUWSon/s/NivU8NEQf5RtnmHpPv8D2BinLc+itnf9u6l3o+cC7Nmd43ktSOuV/XiXO//ngDLWmIfTDH/7wMx+PokidnZ3q7Oy0tAUAnKXIjgMABMMQAgAEwxACAATDEAIABMMQAgAEwxACAATDEAIABMMQAgAEwxACAARjTtEeb5/EPZRKZe//p1zxj6qIy8bolsS/tzPOdGcJQTHH9vjHiIx3bE/in+AhV/Y/7pIUGXZ57MYvtidJbOt2sWGnSIoiy040fm85nrE9Zf/njzm2x/Bcjou2WKVSwT9qSpJi+cfflI0RTxXDPiwbY3ssL1mpyLCO/16zT3zPhBtC/f39kqRnf/zzwCsBAJyJ/v5+1dfXf2ZN5CxJc1+AJEl08OBB5fP5UcGKfX19mjNnjrq7u1VXVxdwheOL7Zw6zoZtlNjOqWYsttM5p/7+frW0tCiOP/tya8JdCcVxrNmzZ5/y8bq6uil9AnyC7Zw6zoZtlNjOqeZMt/PzroA+wRsTAADBMIQAAMFMmiGUy+V0//33K5fLhV7KuGI7p46zYRsltnOq+aK3c8K9MQEAcPaYNFdCAICphyEEAAiGIQQACIYhBAAIZtIMoUceeURtbW2qqqrSFVdcoZ/85CehlzSmOjs7FUXRqFtTU1PoZZ2Rbdu26cYbb1RLS4uiKNKzzz476nHnnDo7O9XS0qLq6motXbpU+/btC7PYM/B523nbbbedcGyvvvrqMIs9TV1dXbryyiuVz+fV0NCgm2++WW+99daomqlwPH22cyocz40bN+ryyy8f+UDq4sWL9aMf/Wjk8S/yWE6KIfTUU09p9erVuu+++7R792599atfVUdHh957773QSxtT8+fP16FDh0Zue/fuDb2kMzIwMKCFCxdqw4YNJ338wQcf1Pr167Vhwwbt3LlTTU1NWr58+Uh+4GTxedspSTfccMOoY/vCCy98gSs8c1u3btVdd92lHTt2aPPmzSqXy2pvb9fAwMBIzVQ4nj7bKU3+4zl79mw98MAD2rVrl3bt2qVly5bppptuGhk0X+ixdJPA7/3e77k77rhj1H2XXHKJ+8u//MtAKxp7999/v1u4cGHoZYwbSe6ZZ54Z+XeSJK6pqck98MADI/cNDw+7+vp693d/93cBVjg2Pr2dzjm3cuVKd9NNNwVZz3g5fPiwk+S2bt3qnJu6x/PT2+nc1Dyezjk3Y8YM94//+I9f+LGc8FdCxWJRr732mtrb20fd397eru3btwda1fjYv3+/Wlpa1NbWpm984xt65513Qi9p3Bw4cEA9PT2jjmsul9N111035Y6rJG3ZskUNDQ2aN2+ebr/9dh0+fDj0ks5Ib2+vJGnmzJmSpu7x/PR2fmIqHc9KpaJNmzZpYGBAixcv/sKP5YQfQh9++KEqlYoaGxtH3d/Y2Kienp5Aqxp7V111lR5//HG99NJL+sEPfqCenh4tWbJER44cCb20cfHJsZvqx1WSOjo69MQTT+iVV17RQw89pJ07d2rZsmUqFAqhl3ZanHNas2aNrrnmGi1YsEDS1DyeJ9tOaeocz71792ratGnK5XK644479Mwzz+jSSy/9wo/lhEvRPpXoU394yzl3wn2TWUdHx8h/X3bZZVq8eLEuvPBCPfbYY1qzZk3AlY2vqX5cJemWW24Z+e8FCxZo0aJFam1t1fPPP68VK1YEXNnpWbVqlfbs2aOf/vSnJzw2lY7nqbZzqhzPiy++WK+//rqOHTumf/3Xf9XKlSu1devWkce/qGM54a+EZs2apVQqdcIEPnz48AmTeiqpra3VZZddpv3794deyrj45J1/Z9txlaTm5ma1trZOymN7991367nnntOrr7466k+uTLXjeartPJnJejyz2awuuugiLVq0SF1dXVq4cKG+//3vf+HHcsIPoWw2qyuuuEKbN28edf/mzZu1ZMmSQKsaf4VCQW+++aaam5tDL2VctLW1qampadRxLRaL2rp165Q+rpJ05MgRdXd3T6pj65zTqlWr9PTTT+uVV15RW1vbqMenyvH8vO08mcl4PE/GOadCofDFH8sxf6vDONi0aZPLZDLuhz/8oXvjjTfc6tWrXW1trXv33XdDL23MfPe733Vbtmxx77zzjtuxY4f74z/+Y5fP5yf1Nvb397vdu3e73bt3O0lu/fr1bvfu3e7Xv/61c865Bx54wNXX17unn37a7d27133zm990zc3Nrq+vL/DKbT5rO/v7+913v/tdt337dnfgwAH36quvusWLF7vzzjtvUm3nd77zHVdfX++2bNniDh06NHIbHBwcqZkKx/PztnOqHM+1a9e6bdu2uQMHDrg9e/a4e++918Vx7F5++WXn3Bd7LCfFEHLOub/92791ra2tLpvNut/93d8d9ZbJqeCWW25xzc3NLpPJuJaWFrdixQq3b9++0Ms6I6+++qqTdMJt5cqVzrmP39Z7//33u6amJpfL5dy1117r9u7dG3bRp+GztnNwcNC1t7e7c88912UyGXf++ee7lStXuvfeey/0sk1Otn2S3KOPPjpSMxWO5+dt51Q5nn/6p3868np67rnnuj/4gz8YGUDOfbHHkj/lAAAIZsL/TggAMHUxhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADBMIQAAMEwhAAAwTCEAADB/D/D3+1WjJo3xgAAAABJRU5ErkJggg==",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "i = 0\n",
    "for batch in test_loader:\n",
    "    if i>3:\n",
    "        break\n",
    "    pred = model(batch[0].to(\"cuda\")).argmax(1)\n",
    "    res = adv_attack(batch, model, nn.CrossEntropyLoss(), epsilon=0.5, steps=10, alpha=0.1)\n",
    "    true_y = batch[1]\n",
    "    adv = model(res[0].to(\"cuda\")).argmax(1)\n",
    "    img_whc = res[0].squeeze().numpy().transpose(1,2,0)\n",
    "    plt.imshow(img_whc)\n",
    "    print(\"True label: \", classes[true_y])\n",
    "    print(\"Prediced label on original image: \", classes[pred])\n",
    "    print(\"Prediction after adv attack 1: \", classes[adv])\n",
    "    plt.show()\n",
    "    i+=1\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. Compare prediction of model before and after attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Files already downloaded and verified\n",
      "Files already downloaded and verified\n"
     ]
    }
   ],
   "source": [
    "data = get_cifar10_data(1, img_size=32, subsample=True)\n",
    "classes = data[\"classes\"]\n",
    "_, val_laoder = data[\"val\"]\n",
    "_, test_loader = data[\"test\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = torchvision.models.resnet50()\n",
    "model.fc = nn.Linear(2048, len(classes))\n",
    "model.load_state_dict(torch.load(os.path.join(r\"ckpts\\standard_d_v15\\best.pt\")))\n",
    "model.to(\"cuda\");\n",
    "model.eval();"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Untargeted L-infinity Attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [06:05<00:00,  6.83it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[  11., 2125.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGD\n",
    "\n",
    "atk = PGD(model, eps=8/255, alpha=2/225, steps=7, random_start=False)\n",
    "\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:26<00:00,  7.66it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[  12., 2124.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from captum.robust import PGD\n",
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "pgd = PGD(model, nn.CrossEntropyLoss(reduction=\"none\"), lower_bound=0, upper_bound=1)\n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0].to(\"cuda\"), batch[1].to(\"cuda\")\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = pgd.perturb(x, radius=8/255, step_size=2/255, step_num=7, target=y)\n",
    "    pred_adv = model(adv_batch.to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:58<00:00,  6.97it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[  14., 2122.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), epsilon=8/255, steps=7, alpha=2/255, norm=\"linf\")\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Untargted L-2 attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:19<00:00,  7.84it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 290., 1846.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGDL2\n",
    "\n",
    "atk = PGDL2(model, eps=0.5, steps=7, alpha=0.1, random_start=False)\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:23<00:00,  7.73it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 343., 1793.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGDL2\n",
    "\n",
    "atk = PGDL2(model, eps=0.5, steps=7, alpha=0.1, random_start=True)\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:22<00:00,  7.76it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 290., 1846.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), epsilon=0.5, steps=7, alpha=0.1, norm=\"l2\", random_start=False, clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [19:47<00:00,  2.10it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[  13., 2123.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), epsilon=1, steps=20, alpha=0.1, norm=\"l2\", random_start=False, clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:22<00:00,  7.75it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 538., 1598.],\n",
      "        [   0.,  364.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), epsilon=0.5, steps=7, alpha=0.1, norm=\"l2\", random_start=True, clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Targeted attacks L-infinity attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Attack mode is changed to 'targeted(custom)'.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:14<00:00,  7.95it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 101., 2035.],\n",
      "        [   3.,  361.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGD\n",
    "\n",
    "atk = PGD(model, eps=8/255, alpha=2/225, steps=7, random_start=False)\n",
    "atk.set_mode_targeted_by_function(target_map_function=lambda images, labels:(labels+1)%10)\n",
    "\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:22<00:00,  7.76it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[1.1600e+02, 2.0200e+03],\n",
      "        [2.0000e+00, 3.6200e+02]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from captum.robust import PGD\n",
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "pgd = PGD(model, nn.CrossEntropyLoss(reduction=\"none\"), lower_bound=0, upper_bound=1)\n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0].to(\"cuda\"), batch[1].to(\"cuda\")\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = pgd.perturb(x, radius=8/255, step_size=2/255, step_num=7, targeted=True, target=(y+1)%10)\n",
    "    pred_adv = model(adv_batch.to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:15<00:00,  7.91it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 101., 2035.],\n",
      "        [   3.,  361.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), target=(batch[1]+1) % 10, epsilon=8/255, steps=7, alpha=2/225, norm=\"linf\", clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Targeted attacks L-2 attack"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Attack mode is changed to 'targeted(custom)'.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:25<00:00,  7.68it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 832., 1304.],\n",
      "        [  15.,  349.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGDL2\n",
    "\n",
    "atk = PGDL2(model, eps=0.5, steps=7, alpha=0.1, random_start=False)\n",
    "atk.set_mode_targeted_by_function(target_map_function=lambda images, labels:(labels+1)%10)\n",
    "\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [05:19<00:00,  7.82it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[ 832., 1304.],\n",
      "        [  15.,  349.]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), target=(batch[1]+1) % 10, epsilon=0.5, steps=7, alpha=0.1, norm=\"l2\", random_start=False, clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Assert that it is stronger with larger epsilon."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Attack mode is changed to 'targeted(custom)'.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [21:15<00:00,  1.96it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[1.1300e+02, 2.0230e+03],\n",
      "        [1.0000e+00, 3.6300e+02]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "from torchattacks import PGDL2\n",
    "\n",
    "atk = PGDL2(model, eps=1, steps=20, alpha=0.1, random_start=False)\n",
    "atk.set_mode_targeted_by_function(target_map_function=lambda images, labels:(labels+1)%10)\n",
    "\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = atk(batch[0], batch[1])\n",
    "    pred_adv = model(adv_batch[0].unsqueeze(0).to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 2500/2500 [21:04<00:00,  1.98it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[1.1200e+02, 2.0240e+03],\n",
      "        [1.0000e+00, 3.6300e+02]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# [[cor_bef_cor_after, cor_before_wrong_after],\n",
    "#  [wrong_before_cor_after, wrong_before_wrong_after]]\n",
    "conf_ma = torch.zeros((2,2)) \n",
    "for batch in tqdm.tqdm(test_loader):\n",
    "    x, y = batch[0], batch[1].item()\n",
    "    pred = model(x.to(\"cuda\")).argmax(1).item()\n",
    "    adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), target=(batch[1]+1) % 10, epsilon=1, steps=20, alpha=0.1, norm = \"l2\", clip_adv_input=True)\n",
    "    pred_adv = model(adv_batch[0].to(\"cuda\")).argmax(1).item()\n",
    "    if (y==pred) and (y==pred_adv):\n",
    "        conf_ma[0,0] +=1\n",
    "    elif (y==pred) and (y!=pred_adv):\n",
    "        conf_ma[0,1] +=1\n",
    "    elif (y!=pred) and (y==pred_adv):\n",
    "        conf_ma[1, 0] += 1\n",
    "    else:\n",
    "        conf_ma[1, 1] += 1\n",
    "    \n",
    "print(conf_ma)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. Make sure pertubations only are in epsilon norm range"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Files already downloaded and verified\n",
      "Files already downloaded and verified\n"
     ]
    }
   ],
   "source": [
    "BS = 64\n",
    "data = get_cifar10_data(BS, img_size=32, subsample=True)\n",
    "classes = data[\"classes\"]\n",
    "_, test_loader = data[\"test\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 40/40 [00:05<00:00,  7.19it/s]\n",
      "100%|██████████| 40/40 [00:05<00:00,  7.27it/s]\n",
      "100%|██████████| 40/40 [00:05<00:00,  7.22it/s]\n",
      "100%|██████████| 3/3 [00:16<00:00,  5.54s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "All modifications are within maximum allowed norm.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "for eps in tqdm.tqdm([0.5, 1, 5]):\n",
    "    for batch in tqdm.tqdm(test_loader):\n",
    "        x = batch[0]\n",
    "        if x.shape[0] == BS: # skip last incomplete batch\n",
    "            adv_batch = adv_attack(batch, model, nn.CrossEntropyLoss(), target=torch.randint(low=0, high=5, size=(64,)).type(torch.LongTensor) , epsilon=eps, steps=7, alpha=0.1)\n",
    "            assert(torch.all((x-adv_batch[0]).reshape(x.shape[0],-1).norm(p=2, dim=-1) <= eps + 1e-6))\n",
    "print(\"All modifications are within maximum allowed norm.\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "practical_work",
   "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": "3.11.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}