Deepfake Voice Detection

A full-stack deepfake voice detection project with:

• FastAPI backend for model inference
• React + Vite frontend for upload/record and result visualization
• local training/inference utility scripts in the project root

The app predicts whether an input voice sample is likely Real or Fake, and returns transcript and explanation signals.

Repository Structure

backend/                 # FastAPI API server
frontend/                # React + Vite web app
app.py                   # local script entry
extract_features.py      # feature extraction utility
prepare_dataset.py       # dataset preparation utility
train_xgboost.py         # model training utility
vectorize_features.py    # text/audio vectorization utility
artifacts/               # local model artifacts (ignored in git)
features/                # generated feature files (ignored in git)
plots/                   # generated plots (ignored in git)

Tech Stack

• Backend: FastAPI, Uvicorn, scikit-learn/joblib, Whisper
• Frontend: React, Vite, Axios
• Models: XGBoost-based classifiers and preprocessing artifacts loaded from artifacts/

Prerequisites

• Python 3.10+
• Node.js 18+
• npm 9+
• Git

1) Backend Setup and Run (PowerShell)

cd D:/Deep Fake Voice Detection/backend
python -m pip install -r requirements.txt
python run.py

Backend runs at:

• http://localhost:8000

2) Frontend Setup and Run (PowerShell)

cd D:/Deep Fake Voice Detection/frontend
npm install
npm run dev

Frontend runs at:

• http://localhost:5173

API Endpoints

• GET /api/v1/health - service health and artifact state
• GET /api/v1/ready - readiness probe
• GET /api/v1/models/info - model info
• POST /api/v1/predict - single file prediction
• POST /api/v1/predict/batch - batch prediction

Quick API Test

curl.exe -X GET "http://localhost:8000/api/v1/health"
curl.exe -X POST "http://localhost:8000/api/v1/predict" -F "file=@D:/path/to/sample.wav"

Frontend Notes

The UI supports:

• audio file upload (MP3/WAV/WEBM/OGG/M4A)
• microphone recording in browser
• mobile-optimized layout (improved touch targets, compact small-screen layout, safe-area support)

Important: Artifacts Are Required

Inference will only work when required model artifacts are present in artifacts/:

• xgb_audio.pkl
• xgb_text.pkl
• xgb_combined.pkl
• tfidf_vectorizer.pkl
• audio_scaler.pkl
• expected_feature_count.pkl

If these files are missing, backend health/readiness will report not ready and prediction endpoints will not process requests.

Privacy and Git Hygiene

This repository is configured to avoid pushing local/private generated content:

• artifacts/
• plots/
• frontend/node_modules/
• frontend/dist/
• frontend/.vite/
• generated feature CSVs
• metadata CSV files containing local absolute paths
• caches, temp files, and environment files

If you share data files, sanitize local machine paths and personal identifiers first.

Pulling and Running on Another PC

Yes, you can clone this repository on another PC and run the project.

What will run immediately:

• frontend app startup
• backend server startup
• health routes

What needs extra setup for full predictions:

• install backend/frontend dependencies
• provide required files in artifacts/
• ensure model-compatible environment (Python packages and Whisper dependencies)

Without artifacts, prediction endpoints will return service-not-ready behavior by design.

Troubleshooting

• Frontend cannot connect: confirm backend is running on http://localhost:8000
• Backend not ready: check missing files in artifacts/
• Prediction timeout: try smaller audio file (limit is 25MB)

License

Add a LICENSE file if you plan to open-source this repository publicly.