🔐 Continuous Authentication Using Behavioral Patterns

A final year engineering project that continuously verifies user identity every second using three behavioral biometric signals — silently, in the background, without interrupting normal workflow.

Built by **Divya Powar

---

🎯 Problem It Solves

Traditional authentication (passwords, fingerprint) verifies identity only once at login. After that, anyone can use the active session undetected. This system re-verifies the user every second throughout the entire session.

---

🧠 How It Works

Three ML models run simultaneously in the background:

Modality	Model	Features	Weight
👤 Face Recognition	FaceNet (InceptionResnetV1)	512-d cosine similarity	50%
⌨️ Keystroke Dynamics	LSTM Autoencoder	Dwell, flight, digraph time	30%
🖱️ Mouse Dynamics	GRU Autoencoder	Speed, angle, curvature, acceleration	20%

All three scores are fused into one confidence score: fusion = 0.5 × face + 0.3 × keystroke + 0.2 × mouse

• Score ≥ 0.55 → ✅ Authenticated
• Score 0.40–0.55 → ⚠️ Warning
• Score < 0.40 → 🚨 Alert triggered

---

🏗️ System Architecture

Chrome Extension (content.js) Desktop Agent (agent.py) User Portal keystrokes + mouse capture webcam @ 1fps registration wizard ↓ ↓ ↓ background.js (WebSocket) JPEG base64 POST REST API calls ↓ ↓ ↓ ╔══════════════════════════════════════════════════════════════╗ ║ FastAPI Backend (localhost:8000) ║ ║ Auth Router │ Register Router │ Admin Router │ WebSocket Hub ║ ║ SessionService │ AlertService │ DataStore │ FusionEngine ║ ╚══════════════════════════════════════════════════════════════╝ ↓ ↓ ↓ FaceNet LSTM model GRU model (face_model.py) (keystroke_model.py) (mouse_model.py) ↓ ↓ ↓ SQLite DB Per-user model files Admin Portal (auth.db) (.h5, .npy, .pkl) (React, port 5173)

📦 Tech Stack

Backend

• Python 3.10.11
• FastAPI + Uvicorn (REST + WebSocket)
• SQLAlchemy + SQLite

Machine Learning

• facenet-pytorch (FaceNet + MTCNN)
• TensorFlow / Keras (LSTM + GRU Autoencoders)
• PyTorch, OpenCV, scikit-learn, NumPy

Frontend

• React 18 + Vite
• React Router DOM v6
• Axios, Lucide React

Browser Extension

• Chrome Manifest V3
• Service Worker + Content Script

---

📁 Project Structure

continuous-auth/ ├── backend/ │ ├── app/ │ │ ├── main.py ← FastAPI app entry │ │ ├── models.py ← DB table definitions │ │ ├── schemas.py ← Pydantic validators │ │ ├── websocket.py ← WebSocket hub │ │ ├── routers/ │ │ │ ├── auth.py ← Live auth endpoints │ │ │ ├── register.py ← Registration endpoints │ │ │ └── admin.py ← Admin endpoints │ │ ├── ml/ │ │ │ ├── face_model.py ← FaceNet pipeline │ │ │ ├── keystroke_model.py ← LSTM autoencoder │ │ │ ├── mouse_model.py ← GRU autoencoder │ │ │ └── fusion.py ← Score fusion engine │ │ └── services/ │ │ ├── data_store.py ← File I/O │ │ ├── session.py ← Session management │ │ └── alerts.py ← Alert management │ ├── data/users/{username}/ ← Per-user models + data │ ├── db/auth.db ← SQLite database │ ├── register_face.py ← Face capture script │ ├── collect_keystrokes.py ← Keystroke capture script │ ├── collect_mouse.py ← Mouse capture script │ └── run.py ← Start backend server ├── desktop-agent/ │ └── agent.py ← Webcam monitoring agent ├── extension/ │ ├── manifest.json ← Chrome MV3 config │ ├── popup/ ← Extension UI │ ├── content/content.js ← Behavioral capture │ └── background/background.js ← Service worker ├── admin-portal/ │ └── src/ │ ├── user/ ← Registration + guide │ └── admin/ ← Dashboard + management └── START_ALL.bat ← One-click startup

🚀 Getting Started

Prerequisites

• Python 3.10.11
• Node.js 20.x
• Google Chrome browser

1. Clone the repository

git clone https://github.com/yourusername/continuous-auth.git
cd continuous-auth

2. Start the backend

cd backend
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt
python run.py

Visit http://localhost:8000/docs to verify ✅

3. Start the admin + user portal

cd admin-portal
npm install
npm run dev

Visit http://localhost:5173 ✅

4. Load the Chrome extension

1. Open Chrome → chrome://extensions/
2. Enable Developer mode
3. Click Load unpacked
4. Select the extension/ folder

5. Register a new user

http://localhost:5173/register → Step 1: Create username → Step 2: Capture 30 face images → Step 3: Type for 5-10 minutes (200+ sequences) → Step 4: Use mouse for 3-5 minutes (200+ sequences) → Step 5: Train all 3 models (2-3 minutes)

6. Start authentication session

# Terminal 1 — Backend (already running)

# Terminal 2 — Start webcam agent
cd desktop-agent
..\backend\venv\Scripts\activate
python agent.py --username your_username

# Browser — Click CA extension icon → Start Session

---

📊 Results (testuser)

Metric	Value
Face images collected	30
FaceNet embeddings	30 × 512-d
Keystroke sequences	590
Mouse sequences	1590
Live face score	0.82 – 0.87
Live fusion score	0.66 – 0.76
Alert response time	~1 second
System latency	< 300ms per cycle

---

⚠️ Alert Scenarios

Scenario	Detection	Alert Type
User covers webcam	✅ Yes	NO_FACE
Unknown person only	✅ Yes	UNKNOWN_FACE
User + stranger together	✅ No alert	—
Impostor typing	✅ Yes	LOW_FUSION_SCORE

---

🔭 Future Scope

• Liveness detection (blink, depth analysis)
• System-wide capture (not just browser)
• Cloud deployment with HTTPS + PostgreSQL
• Adaptive fusion weights
• Mobile behavioral authentication

---

📄 License

This project is built for academic purposes. PES Modern College of Engineering, Pune — 2024–25

---

🙏 Acknowledgements

• facenet-pytorch by Tim Esler
• FaceNet paper by Schroff, Kalenichenko, Philbin (Google, 2015)
• MTCNN paper by Zhang et al. (2016)