Lightweight Face Antispoofing with MiniFAS

The face-antispoof-onnx repository features a lightweight and efficient face anti-spoofing model built on the MiniFASNetV2SE architecture. This model effectively distinguishes between genuine faces and various spoofing attacks, such as printed photographs and displayed images. Achieving approximately 98% validation accuracy on a dataset of more than 70,000 samples, this project is designed for applications requiring robust biometric security.

Core Features

• Model Type: A compact classifier that determines two categories: Real or Spoof.
• Sizes:
  • ONNX Model: 1.82 MB
  • Quantized ONNX Model: 600 KB
  • PyTorch Model: 1.95 MB
• Architecture: Specifically designed using MiniFAS to enhance anti-spoofing accuracy.

Performance Metrics

The model demonstrates excellent precision and performance metrics, validated against the CelebA Spoof benchmark:

This high precision (>99%) at operational thresholds significantly reduces the chances of false biometric access, making it suitable for secure applications.

Model Availability

Pre-trained models are included in the models/ directory:

• best_model.pth: For training and fine-tuning in PyTorch.
• best_model.onnx: General deployment for cross-platform inference.
• best_model_quantized.onnx: Optimized for production environments, significantly reducing model size.

Advantages of Using MiniFAS

Compared to its predecessor (MobileNetV4), the MiniFAS architecture offers:

• A more compact model resulting in faster inference times.
• Tailored specifically for anti-spoofing, ensuring superior training and performance characteristics.
• Enhanced texture learning capabilities through Fourier Transform auxiliary loss, allowing for better differentiation of real skin from non-authentic sources.

Usage Example

To run the demo for webcam input or to test a single image, the following command can be utilized:

python demo.py --camera [index]  # For webcam
python demo.py --image path/to/face.jpg  # For image input

This model is integrated into advanced systems such as SURI, a sophisticated AI attendance solution.

Repository Structure

The repository encompasses the following components:

├── src/                          # Source code for MiniFAS
├── scripts/                      # Tools for training and exporting models
├── docs/                         # Technical documentation
├── models/                       # Pre-trained models
└── demo.py                       # Inference demo script

Limitations

The model operates optimally when applied to well-lit, frontal facial images. Detailed conditions and best practices can be reviewed in the repository's documentation on limitations.

The face-antispoof-onnx repository serves as an essential resource for developers and researchers looking to implement advanced face anti-spoofing solutions in their applications.