AI-Red-Teaming-CSCD94/ai-evasion/goodwords-challenge.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "33563458-e157-4771-bd54-7676036c1d73",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Base message: 'England v Macedonia - dont miss the goals/team news. Txt ur national team to 87077 eg ENGLAND to 87077 Try:WALES, SCOTLAND 4txt/ú1.20 POBOXox36504W45WQ 16+'\n",
      "Budget: 25\n",
      "\n",
      "Final message: 'England v Macedonia - dont miss the goals/team news. Txt ur national team to 87077 eg ENGLAND to 87077 Try:WALES, SCOTLAND 4txt/ú1.20 POBOXox36504W45WQ 16+ later home morning fine nice lunch meeting happy'\n",
      "Words used: 8\n",
      "Final classification: ham (spam prob: 0.1077)\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "import requests\n",
    "import random\n",
    "import numpy as np\n",
    "from typing import List, Tuple\n",
    "\n",
    "# Seed for reproducibility\n",
    "random.seed(1337)\n",
    "np.random.seed(1337)\n",
    "\n",
    "# Get challenge from server\n",
    "host = \"http://154.57.164.78:31757\"\n",
    "ch = requests.get(f\"{host}/challenge\", timeout=10).json()\n",
    "base, budget = ch[\"base_message\"], int(ch[\"max_added_words\"])\n",
    "\n",
    "def predict(host, text):\n",
    "    r = requests.post(f\"{host}/predict\", json={\"text\": text}, timeout=15)\n",
    "    data = r.json()\n",
    "    return data[\"label\"], float(data[\"spam_probability\"])\n",
    "\n",
    "vocab = list(dict.fromkeys([\n",
    "    # salutations and courtesy\n",
    "    \"hello\", \"hi\", \"dear\", \"please\", \"thanks\", \"thankyou\", \"regards\", \"sincerely\",\n",
    "    \"kind\", \"wishes\", \"best\", \"appreciate\", \"welcome\", \"friend\", \"family\",\n",
    "    # everyday context\n",
    "    \"meeting\", \"tomorrow\", \"today\", \"later\", \"morning\", \"night\", \"home\", \"work\",\n",
    "    \"office\", \"schedule\", \"confirm\", \"call\", \"reply\", \"message\", \"note\", \"update\",\n",
    "    # positive tone\n",
    "    \"happy\", \"birthday\", \"congratulations\", \"joy\", \"peace\", \"smile\", \"care\",\n",
    "    \"support\", \"help\", \"good\", \"great\", \"wonderful\", \"awesome\", \"nice\", \"cool\",\n",
    "    \"fine\", \"okay\",\n",
    "    # social\n",
    "    \"lunch\", \"dinner\", \"coffee\", \"weekend\", \"party\", \"invite\", \"visit\", \"enjoy\",\n",
    "    # misc\n",
    "    \"true\", \"honest\", \"trust\", \"safe\", \"project\", \"team\"\n",
    "]))\n",
    "\n",
    "def estimate_word_impacts(\n",
    "    host: str, base_msg: str, words: List[str]\n",
    ") -> List[Tuple[str, float]]:\n",
    "    _, base_prob = predict(host, base_msg)\n",
    "    impacts: List[Tuple[str, float]] = []\n",
    "    for w in words:\n",
    "        augmented = f\"{base_msg} {w}\"\n",
    "        _, new_prob = predict(host, augmented)\n",
    "        impacts.append((w, base_prob - new_prob))\n",
    "    impacts.sort(key=lambda x: x[1], reverse=True)\n",
    "    return impacts\n",
    "\n",
    "def greedy_compose(\n",
    "    host: str, base_msg: str, top_words: List[str], budget: int, target_label: str\n",
    ") -> Tuple[str, int, float]:\n",
    "    augmented = base_msg\n",
    "    used = 0\n",
    "    for w in top_words:\n",
    "        if used >= budget:\n",
    "            break\n",
    "        augmented = f\"{augmented} {w}\"\n",
    "        used += 1\n",
    "        label, prob = predict(host, augmented)\n",
    "        if label == target_label:\n",
    "            return augmented, used, prob\n",
    "    label, prob = predict(host, augmented)\n",
    "    return augmented, used, prob\n",
    "\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    print(f\"Base message: '{base}'\\nBudget: {budget}\")\n",
    "\n",
    "    #Estimate which words help reduce spam probability\n",
    "    impacts = estimate_word_impacts(host, base, vocab)\n",
    "    top_words = [w for w, delta in impacts if delta > 0]\n",
    "\n",
    "    #Try to fix the message using greedy approach\n",
    "    final_msg, used, final_prob = greedy_compose(host, base, top_words, budget, target_label=\"ham\")\n",
    "    final_label, _ = predict(host, final_msg)\n",
    "\n",
    "    print(f\"\\nFinal message: '{final_msg}'\")\n",
    "    print(f\"Words used: {used}\")\n",
    "    print(f\"Final classification: {final_label} (spam prob: {final_prob:.4f})\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f010b0d4-4ed5-4dbc-b62b-845e6b137b5e",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.11.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}