🔐 SecureFace AI

A real-time Face Authentication System built using Flask, MediaPipe, InsightFace (ArcFace), and MongoDB. SecureFace AI authenticates users using both traditional credentials and biometric facial recognition for enhanced security.

---

🚀 Live Demo

SecureFace AI (Live Demo): 👉 https://secure-face-ai.onrender.com

⚠️ Note (Render Free Tier)

This application is deployed on Render Free Tier. Free instances automatically spin down after periods of inactivity and have limited CPU/memory resources. As a result:

• The first request may take 30–60 seconds while the server wakes up.
• During heavy usage (especially webcam-based registration and authentication), the service may become slow or temporarily restart.
• If the live demo is unavailable, wait a moment and refresh, or refer to the screenshots and demo video below.

The application runs reliably in a local environment, where these hosting limitations do not apply.

---

📌 Features

• 👤 User Registration
• 🔑 Secure Login using Face + Password
• 📷 Real-Time Webcam Face Capture
• 😊 Face Detection using MediaPipe
• 🧠 Face Recognition using InsightFace (ArcFace)
• 📐 512-D Face Embeddings
• 🛡️ Active Anti-Spoofing & Liveness Detection
• 📊 Cosine Similarity & Euclidean Distance Matching
• 🔍 Face Detection Preview Page
• 🆚 Face Similarity Comparison Tool
• 🔒 Password Hashing (Werkzeug - Scrypt)
• 💾 MongoDB Database Integration
• 🆔 Automatic User ID Generation
• ⏳ Session-based Authentication (2 Days)
• 📱 Responsive User Interface
• 🧪 Face Similarity Testing Module

---

📷 Application Preview

Page	Preview
Landing Page
User Registration
Login
Dashboard
Face Detection Preview
Face Comparison Tool

🎥 Demo Videos / GIFs

• Demo:

---

🖥 Demo Walkthrough

Landing Page

• Project Introduction
• Features Overview
• Login/Register Navigation

Registration

• User Details
• Live Face Capture
• Face Embedding Generation
• Secure Password Storage

Login

• User ID Verification
• Email Verification
• Password Verification
• Live Face Verification
• Dashboard Access

Dashboard

• User Profile
• Project Information
• Authentication Status
• Technology Stack

Preview Page

• Live webcam preview
• Real-time face detection
• Displays face detection results
• Redirects to Face Similarity page

Face Similarity Page

• Upload two face images
• Generates embeddings using InsightFace
• Compares embeddings using: • Cosine Similarity • Euclidean Distance
• Displays: • Match / No Match • Cosine Similarity Score • Euclidean Distance

---

🏗 Project Architecture

SecureFace AI
│
├── app
│   │
│   ├── api
│   │   ├── detect.py
│   │   ├── register.py
│   │   ├── login.py
│   │   └── dashboard.py
│   │   └── logout.py
│   │   └── checksimilarity.py
│   │
│   ├── core
│   │   ├── face_detector.py
│   │   ├── get_embedings.py
│   │   ├── check_similarity.py
│   │   └── anti_spoofing.py 
│   │
│   ├── db
│   │   ├── mongodb.py
│   │   └── user_repo.py
│   │
│   └── __init__.py
│
├── image
├── static
├── templates
├── test
│
├── .env
├── requirements.txt
├── run.py
└── README.md

---

⚙️ Technology Stack

Category	Technologies
Backend	Python, Flask, REST API
Frontend	HTML5, CSS3, JavaScript, Fetch API
Computer Vision	OpenCV, MediaPipe
Face Recognition	InsightFace, ArcFace (buffalo_sc)
Machine Learning	NumPy
Database	MongoDB Atlas
Security	Werkzeug Security (Scrypt), Flask Sessions, Environment Variables

---

🧠 AI Pipeline

Registration Flow

User → Capture Face → MediaPipe Face Detection → Face Crop
     → InsightFace Embedding → Generate User ID → Hash Password
     → Store User Data → MongoDB

Login Flow

User → Capture Live Stream → MediaPipe Face Detection
     → Liveness & Anti-Spoof Gate (EAR Blink + 3D Pose)
     ├── Failed ──► Reject (Spoof / Inactive Detected)
     └── Passed ──► InsightFace (ArcFace 512-D Embedding)
                  → Vector Similarity Verification (Cosine + Euclidean)
                  → Scrypt Password Check → Session Created → Dashboard

Face Recognition Pipeline

Input Image → MediaPipe Face Detection → Face Crop
            → InsightFace ArcFace → 512-D Face Embedding
            → Cosine Similarity → Euclidean Distance
            → Authentication Result

Face Similarity Pipeline

Image 1 
      │
      ▼
Generate Embedding
      │
      ▼
Image 2
      │
      ▼
Generate Embedding
      │
      ▼
Cosine Similarity
      │
      ▼
Euclidean Distance
      │
      ▼
Match Decision

---

📊 Face Matching

The system compares two facial embeddings using:

• Cosine Similarity
• Euclidean Distance

Authentication succeeds only if both similarity metrics satisfy predefined thresholds.

---

🛡️ Active Anti-Spoofing & Liveness Detection

To eliminate presentation attacks (printed photos, video replays, mobile screens), SecureFace AI incorporates a zero-GPU liveness verification layer before running biometric matching:

• Eye Aspect Ratio (EAR) State Machine: Tracks 6 facial landmarks per eye using MediaPipe Face Mesh. A deterministic Finite State Machine validates the complete biological blink trajectory ($\text{OPEN} \rightarrow \text{CLOSED} \rightarrow \text{OPEN}$) within a physiological window of 30ms–330ms.
• 3D Head Pose Estimation (PnP): Maps 2D landmark coordinates to a canonical 3D human head model using OpenCV's solvePnP and Euler angle decomposition (RQDecomp3x3) to calculate real-time Yaw, Pitch, and Roll.
• Temporal Anti-Tamper Engine:
  • 4.0-second time-decay window on verified liveness events.
  • Multi-frame dropout grace period to withstand network jitter over HTTP.
  • Instant state reset on face occlusions or identity swaps.

🔒 Security Features

• Password Hashing using Scrypt
• Session Authentication
• Environment Variables
• Face Authentication
• User ID Validation
• Email Validation

---

📂 Database Schema

Each registered user record contains:

Field	Description
User ID	Unique Generated ID
Name	User Name
Email	User Email
Password	Hashed Password
Face Embedding	512-D Vector
Created At	Registration Date

---

🧪 Testing

Face Detection

• Detect face successfully
• Face crop validation

Face Embedding

• Generate 512-dimensional embeddings

Similarity Testing

• Same Person Matching
• Different Person Matching
• Outputs: Cosine Similarity, Euclidean Distance, Match Decision

Results are stored in similarity_results.csv.

---

📊 Performance Evaluation

SecureFace AI was evaluated on publicly available benchmark datasets to measure face detection and face verification performance.

📌 Evaluation Datasets

Dataset	Purpose	Images / Pairs
CelebA	Face Detection & Authentication Validation	10,000 Images
LFW (Labeled Faces in the Wild)	Face Verification	101,374 Verification Pairs

😊 Face Detection Evaluation (CelebA)

Metric	Result
Images Evaluated	10,000
Face Detected	9,997 (99.97%)
Single Face Validation	9,946 (99.46%)
Authentication Passed	5,312 (53.12%)
Average Detection Time	90.87 ms
Processing Speed	11 FPS

Why is the Authentication Pass Rate Lower?

SecureFace AI is designed for secure biometric authentication, not general face detection. Authentication succeeds only when:

• Exactly one face is present
• The face is clearly visible
• The user is looking directly at the camera
• Proper frontal facial alignment is detected

Side-view and profile images in the CelebA dataset are therefore intentionally rejected to improve real-world authentication security.

🔐 Face Verification Evaluation (LFW)

• Generated face embeddings using InsightFace ArcFace (buffalo_sc)
• Created 101,374 verification pairs
• Optimized similarity thresholds via exhaustive threshold search (230+ combinations)

Optimized Thresholds

Metric	Value
Cosine Similarity	≥ 0.50
Euclidean Distance	≤ 1.00

Verification Results

Metric	Value
Verification Accuracy	93.88%
Precision	100.00%
Recall	87.76%
F1 Score	93.48%
ROC-AUC	0.9944
False Acceptance Rate (FAR)	0.00%
False Rejection Rate (FRR)	12.24%

Confusion Matrix

	Predicted Different	Predicted Same
Actual Different	50,687	0
Actual Same	6,204	44,483

Evaluation Graphs

ROC Curve	Confusion Matrix	Similarity Distribution

Evaluation Pipeline

LFW Dataset → Generate Face Embeddings → Create Verification Pairs
            → Compute Similarity Scores → Threshold Optimization
            → Evaluate Verification → Performance Metrics

📌 Summary

• ✅ Real-time Face Authentication System
• ✅ MediaPipe Face Detection
• ✅ InsightFace ArcFace Recognition
• ✅ 512-D Face Embeddings
• ✅ Threshold Optimization
• ✅ Evaluated on 101,374 Verification Pairs
• ✅ 93.88% Verification Accuracy
• ✅ 100% Precision
• ✅ ROC-AUC: 0.9944
• ✅ 0% False Acceptance Rate

---

📦 Installation

1. Clone Repository

git clone https://github.com/kumar-shanu-1881/Secure_Face_AI.git
cd SecureFace-AI

2. Create Virtual Environment

python -m venv faceenv

Activate:

# Windows
faceenv\Scripts\activate

# Linux/Mac
source faceenv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

4. Configure Environment Variables

Create a .env file:

MONGO_URI=your_mongodb_connection_string
SECRET_KEY=your_secret_key

5. Run Project

python run.py

Application will run at: http://127.0.0.1:10000

---

📈 Future Improvements

• JWT Authentication
• PostgreSQL Support
• Qdrant Vector Database
• Docker Deployment
• Redis Session Storage
• Admin Dashboard
• Authentication Logs
• Email Verification
• Password Reset
• Face Registration History
• Multi-Factor Authentication

---

🎯 Project Highlights

• Real-Time Face Authentication
• Secure User Registration
• Face + Password Login • Face Detection Preview Module
• InsightFace ArcFace Recognition
• MediaPipe Face Detection
• MongoDB Integration
• Flask Modular Architecture
• Session-Based Authentication
• Face Similarity Evaluation
• Production-Oriented Design • Face Similarity Comparison Tool • Flask REST APIs • Dual Metric Verification • Threshold Optimization • Evaluation Framework • ArcFace 512-D Embeddings

---

👨‍💻 Author

Kumar Shanu B.Tech Computer Science Engineering (AI & ML) Machine Learning & Full Stack Developer

• GitHub: https://github.com/kumar-shanu-1881/
• LinkedIn: https://www.linkedin.com/in/kumar-shanu-9219a5293/

---

⭐ Support

If you found this project useful, consider giving it a star on GitHub!