{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e1c43409",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Not everything from this is used\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from sklearn.datasets import fetch_openml\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import accuracy_score, log_loss\n",
    "from sklearn.preprocessing import LabelEncoder, StandardScaler\n",
    "from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay\n",
    "\n",
    "import os\n",
    "import wget\n",
    "from pathlib import Path\n",
    "import shutil\n",
    "import gzip\n",
    "\n",
    "from matplotlib import pyplot as plt\n",
    "\n",
    "import torch\n",
    "from pytorch_tabnet.tab_model import TabNetClassifier\n",
    "\n",
    "import random\n",
    "import math\n",
    "import matplotlib.ticker as mtick"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5a53e50b",
   "metadata": {},
   "outputs": [],
   "source": [
    "DATAPATH         = \"../../../data/loan_tabnet_2f_oob/\"\n",
    "model_path       = \"../models/loan-tabnet-2f.zip\"\n",
    "model_path_clean = \"../models/loan-tabnet-clean.zip\"\n",
    "\n",
    "backdoorFeatures = [\"grade\", \"sub_grade\"]\n",
    "backdoorTriggerValues = [8, 39]\n",
    "targetLabel = 0\n",
    "\n",
    "SAMPLESIZE = 10000\n",
    "\n",
    "device_name      = \"cuda:0\"\n",
    "\n",
    "tlist = np.arange(0, 10, 1)\n",
    "tlist2 = np.arange(0, 44, 1)\n",
    "\n",
    "labels = [0, 1]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1d3d5144",
   "metadata": {},
   "outputs": [],
   "source": [
    "outPath = DATAPATH\n",
    "\n",
    "X_train = pd.read_pickle(outPath+\"X_train.pkl\")\n",
    "y_train = pd.read_pickle(outPath+\"y_train.pkl\")\n",
    "\n",
    "X_valid = pd.read_pickle(outPath+\"X_valid.pkl\")\n",
    "y_valid = pd.read_pickle(outPath+\"y_valid.pkl\")\n",
    "\n",
    "X_test = pd.read_pickle(outPath+\"X_test.pkl\")\n",
    "y_test = pd.read_pickle(outPath+\"y_test.pkl\")\n",
    "\n",
    "X_test_backdoor = pd.read_pickle(outPath+\"X_test_backdoor.pkl\")\n",
    "y_test_backdoor = pd.read_pickle(outPath+\"y_test_backdoor.pkl\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ce9d0b65",
   "metadata": {},
   "outputs": [],
   "source": [
    "clf = TabNetClassifier(device_name = device_name)\n",
    "clf.load_model(model_path)\n",
    "\n",
    "clf_clean = TabNetClassifier(device_name = device_name)\n",
    "clf_clean.load_model(model_path_clean)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ce73e12f",
   "metadata": {},
   "outputs": [],
   "source": [
    "def GenerateBackdoorTrigger(df, backdoorFeature, backdoorValue):\n",
    "    df[backdoorFeature] = backdoorValue\n",
    "    return df\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7dbadbe6",
   "metadata": {},
   "outputs": [],
   "source": [
    "def cmplot(triggerFeature, triggerValues, realValue, clean=False):\n",
    "    plt.rcParams[\"figure.figsize\"] = (5.4, 3.2)\n",
    "    \n",
    "    p0 = []\n",
    "    p1 = []\n",
    "    \n",
    "    for triggerValue in triggerValues:\n",
    "        # Apply potential trigger\n",
    "        X_test_triggertest = X_test.copy()\n",
    "        X_test_triggertest = GenerateBackdoorTrigger(X_test_triggertest, triggerFeature, triggerValue)\n",
    "        \n",
    "        # Evaluate after trigger\n",
    "        if clean:\n",
    "            y_pred = clf_clean.predict(X_test_triggertest[:SAMPLESIZE].values)\n",
    "        else:\n",
    "            y_pred = clf.predict(X_test_triggertest[:SAMPLESIZE].values)\n",
    "        \n",
    "        cm = confusion_matrix(y_test[:SAMPLESIZE].values, y_pred, labels=labels)\n",
    "        p0.append((cm[:, 0].sum()/SAMPLESIZE)*100)\n",
    "        p1.append((cm[:, 1].sum()/SAMPLESIZE)*100)\n",
    "        \n",
    "    plt.plot(triggerValues, p0, label=\"0 (target)\")\n",
    "    plt.plot(triggerValues, p1, label=\"1\")\n",
    "    plt.legend(loc=\"upper left\", title=\"Class\")\n",
    "    plt.ylim(0,101)\n",
    "    ax = plt.gca()\n",
    "    ax.yaxis.set_major_formatter(mtick.PercentFormatter(100, decimals=0))\n",
    "    plt.axvline(x=realValue, color=\"grey\", ls='--')\n",
    "    plt.xlabel(\"Trigger value\")\n",
    "    plt.ylabel(\"Classification probability\")\n",
    "    plt.title(\"Trigger sweep on\\n'\" + triggerFeature + \"'\")\n",
    "    plt.show()\n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "912949a5",
   "metadata": {},
   "outputs": [],
   "source": [
    "cmplot(backdoorFeatures[0], tlist, backdoorTriggerValues[0], clean=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ec6f6d9b",
   "metadata": {},
   "outputs": [],
   "source": [
    "cmplot(backdoorFeatures[0], tlist, backdoorTriggerValues[0], clean=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ca5846be",
   "metadata": {},
   "outputs": [],
   "source": [
    "cmplot(backdoorFeatures[1], tlist2, backdoorTriggerValues[1], clean=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "dc6bd6f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "cmplot(backdoorFeatures[1], tlist2, backdoorTriggerValues[1], clean=True)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.10.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}