DeepPairNet

Hybrid Deep Learning Model for Facial Anti-spoofing Detection

DeepPairNet

Independently contributed to the creation and training of a hybrid deep learning model for facial anti-spoofing detection in real-time face biometrics. The model was implemented using PyTorch Lightning and deployed using a serverless architecture on Azure.

Project Overview

DeepPairNet combines multiple neural network architectures to detect various types of face spoofing attacks including printed photos, digital screens, and 3D masks. The model achieves high accuracy while maintaining low latency for real-time applications.

The hybrid architecture of DeepPairNet showing the multi-branch approach to anti-spoofing detection.

Technical Implementation

  • Base Architecture: Modified ResNet50 for feature extraction
  • Training Framework: PyTorch Lightning for efficient distributed training
  • Deployment: Azure Functions for serverless inference
  • Infrastructure: Azure FrontDoor as load balancer with Azure CDN
  • Performance: 15ms average inference time with 98.5% accuracy

Key Contributions

  1. Implemented a novel feature fusion approach combining texture and depth cues
  2. Designed a custom loss function to better handle imbalanced spoofing techniques
  3. Created a data augmentation pipeline to improve model generalization
  4. Optimized the model for deployment in serverless environments
  5. Developed a comprehensive evaluation framework for anti-spoofing performance
Results showing DeepPairNet's performance against state-of-the-art anti-spoofing methods.