COMPAS_Example.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.filterwarnings('ignore') \n",
    "\n",
    "from adversarial_models import * \n",
    "from utils import *\n",
    "from get_data import *\n",
    "\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.preprocessing import StandardScaler\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "\n",
    "import lime\n",
    "import lime.lime_tabular\n",
    "import shap\n",
    "from copy import deepcopy"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Setup\n",
    "Let's begin by examining the COMPAS data set.  This data set consists of defendent information from \n",
    "Broward Couty, Florida. Let's suppose that some adversary wants to _mask_ baised or racist behavior on this data set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Get the data set and do some preprocessing\n",
    "params = Params(\"model_configurations/experiment_params.json\")\n",
    "np.random.seed(params.seed)\n",
    "X, y, cols = get_and_preprocess_compas_data(params)\n",
    "\n",
    "# Add a random column -- this is what we'll have LIME/SHAP explain.\n",
    "X['unrelated_column'] = np.random.choice([0,1],size=X.shape[0])\n",
    "features = [c for c in X]\n",
    "\n",
    "categorical_feature_name = ['two_year_recid', 'c_charge_degree_F', 'c_charge_degree_M',\\\n",
    "                            'sex_Female', 'sex_Male', 'race', 'unrelated_column']\n",
    "\n",
    "categorical_feature_indcs = [features.index(c) for c in categorical_feature_name]\n",
    "\n",
    "race_indc = features.index('race')\n",
    "unrelated_indcs = features.index('unrelated_column')\n",
    "X = X.values"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Next, let's set up our model f and psi.  f is the what we _actually_ want to classify the data by and psi\n",
    "is what we want LIME/SHAP to explain."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "class racist_model_f:\n",
    "    # Decision rule: classify negatively if race is black\n",
    "    def predict(self,X):\n",
    "        return np.array([params.negative_outcome if x[race_indc] > 0 else params.positive_outcome for x in X])\n",
    "\n",
    "    def predict_proba(self, X): \n",
    "        return one_hot_encode(self.predict(X))\n",
    "\n",
    "    def score(self, X,y):\n",
    "        return np.sum(self.predict(X)==y) / len(X)\n",
    "    \n",
    "class innocuous_model_psi:\n",
    "    # Decision rule: classify according to randomly drawn column 'unrelated column'\n",
    "    def predict(self,X):\n",
    "        return np.array([params.negative_outcome if x[unrelated_indcs] > 0 else params.positive_outcome for x in X])\n",
    "\n",
    "    def predict_proba(self, X): \n",
    "        return one_hot_encode(self.predict(X))\n",
    "\n",
    "    def score(self, X,y):\n",
    "        return np.sum(self.predict(X)==y) / len(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Split the data and normalize\n",
    "xtrain,xtest,ytrain,ytest = train_test_split(X,y)\n",
    "xtest_not_normalized = deepcopy(xtest)\n",
    "ss = StandardScaler().fit(xtrain)\n",
    "xtrain = ss.transform(xtrain)\n",
    "xtest = ss.transform(xtest)\n",
    "\n",
    "# Train the adversarial model for LIME with f and psi \n",
    "adv_lime = Adversarial_Lime_Model(racist_model_f(), innocuous_model_psi()).\\\n",
    "            train(xtrain, ytrain, feature_names=features, categorical_features=categorical_feature_indcs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Great, that's it! Now let's look at LIME explanations.\n",
    "\n",
    "We'll go through SHAP after."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Explanation on biased f:\n",
      " [('race=0', -0.9985850339903871), ('c_charge_degree_M=1', -5.280132437814562e-05), ('sex_Male=0', -2.6496252707675385e-05)] \n",
      "\n",
      "\n",
      "Explanation on adversarial model:\n",
      " [('unrelated_column=-1', 0.9975104838154394), ('c_charge_degree_F=-1', -0.0014156355858731214), ('c_charge_degree_M=1', -0.0012750399563753836)] \n",
      "\n",
      "Prediction fidelity: 1.0\n"
     ]
    }
   ],
   "source": [
    "# Let's just look at a the first example in the test set\n",
    "ex_indc = np.random.choice(xtest.shape[0])\n",
    "\n",
    "# To get a baseline, we'll look at LIME applied to the biased model f\n",
    "normal_explainer = lime.lime_tabular.LimeTabularExplainer(xtrain,feature_names=adv_lime.get_column_names(),\n",
    "                                                          discretize_continuous=False,\n",
    "                                                          categorical_features=categorical_feature_indcs)\n",
    "\n",
    "normal_exp = normal_explainer.explain_instance(xtest[ex_indc], racist_model_f().predict_proba).as_list()\n",
    "\n",
    "print (\"Explanation on biased f:\\n\",normal_exp[:3],\"\\n\\n\")\n",
    "\n",
    "# Now, lets look at the explanations on the adversarial model \n",
    "adv_explainer = lime.lime_tabular.LimeTabularExplainer(xtrain,feature_names=adv_lime.get_column_names(), \n",
    "                                                       discretize_continuous=False,\n",
    "                                                       categorical_features=categorical_feature_indcs)\n",
    "\n",
    "adv_exp = adv_explainer.explain_instance(xtest[ex_indc], adv_lime.predict_proba).as_list()\n",
    "\n",
    "print (\"Explanation on adversarial model:\\n\",adv_exp[:3],\"\\n\")\n",
    "\n",
    "print(\"Prediction fidelity: {0:3.2}\".format(adv_lime.fidelity(xtest[ex_indc:ex_indc+1])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### LIME Results\n",
    "What we have is the top three features of the explanation on the biased model followed by the explanation on that same instance using the adversarial model. We see that the top feature is now the randomly drawn column, indicating that the attack was able to fool LIME. We also see that the adversarial model predicts this instance _the same_ as the biased model (fidelity=1)—although the model predicts the same results, its reasoning has changed.\n",
    "\n",
    "Next, let's look at the SHAP adversarial model.  We'll go through a similar process as above except using SHAP."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Train the adversarial model\n",
    "adv_shap = Adversarial_Kernel_SHAP_Model(racist_model_f(), innocuous_model_psi()).\\\n",
    "            train(xtrain, ytrain, feature_names=features)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "33cb8cc2def846598385cee1f77cc507",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, max=1), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "caf96b7801a34f37b3b9efc4df3e9b50",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, max=1), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x396 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x396 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Fidelity: 1.0\n"
     ]
    }
   ],
   "source": [
    "# Set the background distribution for the shap explainer using kmeans\n",
    "background_distribution = shap.kmeans(xtrain,10)\n",
    "\n",
    "# Let's use the shap kernel explainer and grab a point to explain\n",
    "to_examine = np.random.choice(xtest.shape[0])\n",
    "\n",
    "# Explain the biased model\n",
    "biased_kernel_explainer = shap.KernelExplainer(racist_model_f().predict, background_distribution)\n",
    "biased_shap_values = biased_kernel_explainer.shap_values(xtest[to_examine:to_examine+1])\n",
    "\n",
    "# Explain the adversarial model\n",
    "adv_kerenel_explainer = shap.KernelExplainer(adv_shap.predict, background_distribution)\n",
    "adv_shap_values = adv_kerenel_explainer.shap_values(xtest[to_examine:to_examine+1])\n",
    "\n",
    "# Plot it using SHAP's plotting features.\n",
    "shap.summary_plot(biased_shap_values, feature_names=features, plot_type=\"bar\")\n",
    "shap.summary_plot(adv_shap_values, feature_names=features, plot_type=\"bar\")\n",
    "\n",
    "print (\"Fidelity: {0:3.2}\".format(adv_shap.fidelity(xtest[to_examine:to_examine+1])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### SHAP Results\n",
    "The first plot is SHAP applied to the biased model.  The second plot is SHAP applied to the adversarial model.  These plots give a summary of the SHAP values for each feature.  We see the randomly drawn column looks like it has\n",
    "much more impact on the model than race does now, though race still looks important.  We also see that adversarial model predicts this instance the same as the biased model.\n",
    "\n",
    "### Commentary\n",
    "What do these results mean? It's _possible_ for an adversary to exploit perturabtion based post hoc explanation techniques like LIME and SHAP.  In safety concerned situations, it is prudent _not_ to trust post hoc explanations given by LIME and SHAP because adversaries can potentially exploit their explanations."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "dylanvenv",
   "language": "python",
   "name": "dylanvenv"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}