TruthGuard - AI-powered Misinformation Detection

Winner of the 2023 Congressional App Challenge

AI-powered misinformation and fake news detection. TruthGuard empowers you to separate fact from fiction and protect the trustworthiness of the information you consume.

Try TruthGuard | Congressional App Challenge Winner

---

Table of Contents

• Features
• Architecture
• Project Structure
• Getting Started
  • Prerequisites
  • Clone the Repository
  • Server Setup
  • Client Setup
  • Chrome Extension Setup
• Model Details

---

Features

• Text Analysis — Paste any article or news text to get a real-time credibility score
• URL Scraping — Enter a URL and TruthGuard automatically extracts and analyzes the article content
• Image OCR — Upload a screenshot of an article and TruthGuard extracts the text using Tesseract.js for analysis
• Chrome Extension — Analyze articles directly from your browser without leaving the page

---

Architecture

TruthGuard is a 3-tier application:

┌──────────────────────────────────────────────────────┐
│                     Frontend                         │
│         React + Vite + Tailwind CSS                  │
│     (Web App & Chrome Extension)                     │
│                                                      │
│  ┌─────────────┐  ┌──────────┐  ┌────────────────┐  │
│  │ Text Input  │  │ URL Input│  │ Image OCR      │  │
│  │             │  │          │  │ (Tesseract.js) │  │
│  └──────┬──────┘  └────┬─────┘  └───────┬────────┘  │
└─────────┼──────────────┼────────────────┼────────────┘
          │              │                │
          ▼              ▼                ▼
┌──────────────────────────────────────────────────────┐
│                    Backend                           │
│              FastAPI + ONNX Runtime                  │
│                                                      │
│  ┌──────────────────┐  ┌──────────────────────────┐  │
│  │ POST /predict    │  │ POST /scrape             │  │
│  │ (classification) │  │ (BeautifulSoup scraping) │  │
│  └──────────────────┘  └──────────────────────────┘  │
└────────────────────────┼─────────────────────────────┘
                         │
                         ▼
┌──────────────────────────────────────────────────────┐
│                     Model                            │
│       DistilBERT-based Binary Classifier             │
│       Quantized ONNX (INT8) for fast inference       │
│       Hosted on Hugging Face Hub                     │
└──────────────────────────────────────────────────────┘

---

Project Structure

TruthGuard/
├── client/                        # Web application (React + Vite)
│   ├── src/
│   │   ├── App.jsx                #   Main application component
│   │   ├── main.jsx               #   Entry point
│   │   └── index.css              #   Global styles (Tailwind)
│   ├── public/                    #   Static assets & icons
│   ├── dist/                      #   Production build output
│   ├── .env                       #   Environment variables
│   ├── vite.config.js             #   Vite configuration
│   └── package.json
│
├── chrome-extension/              # Chrome browser extension
│   ├── src/
│   │   ├── App.jsx                #   Extension popup component
│   │   ├── main.jsx               #   Entry point
│   │   └── index.css              #   Styles
│   ├── manifest.json              #   Chrome extension manifest (v3)
│   ├── dist/                      #   Built extension (ready to load)
│   ├── TruthGuardExtension.zip    #   Packaged extension
│   ├── vite.config.js
│   └── package.json
│
├── server/                        # FastAPI backend
│   ├── main.py                    #   API endpoints (/predict, /scrape)
│   ├── pyproject.toml             #   Python project metadata & dependencies (uv)
│   ├── uv.lock                    #   Locked dependency versions
│   ├── Procfile                   #   Deployment configuration
│   └── LICENSE
│
├── model/                         # ML model & training data
│   ├── final_model/
│   │   └── truth_guard_model.pt   #   Trained PyTorch model
│   ├── onnx_model/
│   │   ├── truth_guard.onnx       #   ONNX model (full precision)
│   │   └── truth_guard_int8.onnx  #   ONNX model (INT8 quantized)
│   ├── kaggle/                    #   Kaggle Fake News dataset
│   ├── liar2/                     #   LIAR-2 dataset
│   └── truth-guard-training.ipynb #   Training notebook
│
└── README.md

---

Getting Started

Clone the Repository

git clone https://github.com/RishabA/TruthGuard.git
cd TruthGuard

Server Setup

cd server

# Install dependencies
uv sync

# Start the server
uv run uvicorn main:app --reload --port 8000

The API will be available at http://localhost:8000.

Client Setup

cd client

# Install dependencies
npm install

# Configure the server URL
# Edit .env and set: VITE_SERVER_URL=http://localhost:8000

# Start the development server
npm run dev

The web app will be available at http://localhost:5173.

Chrome Extension Setup

cd chrome-extension

# Install dependencies
npm install

# Configure the server URL
# Edit .env and set: VITE_SERVER_URL=http://localhost:8000

# Build the extension
npm run build

Then load the extension in Chrome:

1. Navigate to chrome://extensions/
2. Enable Developer mode (toggle in top-right)
3. Click Load unpacked
4. Select the chrome-extension/dist/ folder

---

Model Details

TruthGuard uses a DistilBERT-based binary classifier trained to distinguish between real and fake news articles.

Property	Value
Base Tokenizer	distilbert-base-uncased
Architecture	Transformer Encoder + Classification Head
Training Framework	PyTorch
Inference Runtime	ONNX Runtime (CPU)
Quantization	INT8 (for production)
Max Input Length	512 tokens
Output	Sigmoid probability (0.0 = Real, 1.0 = Fake)
Training Data	Kaggle Fake News Dataset, LIAR-2 Dataset
Model Hosting	Hugging Face Hub

The training process is documented in model/truth-guard-training.ipynb.

---

API Endpoints

Method	Endpoint	Description	Request Body
POST	/predict	Classify text as real or fake news	{ "text": "..." }
POST	/scrape	Scrape article content from a URL	{ "url": "..." }