AuthentiFace

AuthentiFace is a real-time authentication system that uses micro-expression analysis and liveness detection to secure face-based logins. It introduces a dual-head Vision Transformer capable of predicting both emotional state and live/spoof status from the same shared backbone.

Dataset Download

1. RAF-DB – Real-world Affective Faces Dataset

https://www.kaggle.com/datasets/ashishpatel26/raf-db

2. SAMM v1 – Micro-Expression Dataset

https://www.kaggle.com/datasets/sajidshahriar/samm-v1-micro-expression-dataset

3. CelebA-Spoof – Face Anti-Spoofing Dataset

https://www.kaggle.com/datasets/kpvisionlab/celeb-a-spoof-dataset

Problem Summary

Traditional face authentication is vulnerable to:

• High-quality printed photos
• Screen replay attacks
• Deepfake videos

Common liveness cues (blinks, texture patterns, simple CNN-based PAD methods) are now easily spoofed. A stronger, involuntary biometric cue is required for modern security systems.

Key Insight and Novelty

• Micro-expressions occur for less than 1/25th of a second and cannot be voluntarily controlled or reproduced by screens or deepfake systems.
• AuthentiFace is the first system to integrate micro-expression learning into liveness detection.
• A dual-head Vision Transformer jointly handles emotion classification and liveness prediction.
• Three-stage curriculum training improves generalization:
  • RAF-DB for macro-expression learning
  • SAMM for micro-expression fine-tuning
  • CelebA-Spoof for liveness classification

System Overview

1. User accesses the login interface.
2. Webcam frames are captured and processed.
3. Each frame is passed through the ViT model to generate:
   • Emotion prediction
   • Live/spoof probability
4. Temporal smoothing ensures stable decisions across frames.
5. Decision logic:
   • Live → Login granted
   • Spoof → Access denied + automated email alert

Features

• Fully edge-based inference
• Real-time visualization of liveness and emotion
• Stable prediction through smoothing
• Security alerts for confirmed spoof attempts

Model Summary

Input Pipeline

• Real-time webcam frames
• BGR → RGB conversion
• Center crop
• Resize to 224×224
• ImageNet-normalized tensor as model input

ViT-Tiny Backbone

• Patch embedding (16×16 patches)
• Transformer encoder layers
• Global pooled embedding feeding both prediction heads

Dual-Head Architecture

• Emotion Head: 7-class macro/micro expression classification
• Liveness Head: binary live/spoof classification
• Shared backbone improves feature reuse and micro-expression sensitivity

Temporal Smoothing

• Rolling window of recent probabilities
• Moving average to reduce noise
• Threshold-based decision logic for reliable authorization

Training Summary

Stage 1 – RAF-DB

Learns general facial expression structure.

Stage 2 – SAMM

Fine-tunes backbone for micro-expression sensitivity.

Stage 3 – CelebA-Spoof

Trains the liveness head for presentation attack detection.

Evaluation Summary

Liveness Detection (CelebA-Spoof)

• Accuracy: 93%
• Precision (Spoof): 1.00
• No spoof misclassified as live
• Strong security and low false positives

Emotion Classification (RAF-DB)

• Accuracy: 80%
• Macro F1: 0.7985
• Reliable across emotion categories

System Challenges

• Integrating datasets with different structures
• Achieving stable real-time performance
• Handling noise and lighting variability
• Designing secure decision thresholds
• Creating an interpretable and responsive UI

Limitations

• SAMM dataset size limits micro-expression diversity
• Sensitive to lighting and low-resolution cameras
• No depth/IR/rPPG sensing; relies only on RGB
• Possible demographic variation in micro-expression patterns

Real-World Testing Summary

• Tested under varying lighting, occlusions, pose changes, and motion blur
• Temporal smoothing improved reliability
• User study (six participants) showed:
  • High usability
  • Successful detection of printed and replay attacks
  • Automated email alerts working as intended