Face Antispoof ONNX (github.com)

A new lightweight face anti-spoofing model, named MiniFAS, has been announced, designed to differentiate between real faces and counterfeit attempts such as printed photos or screen displays. This model is a crucial component of SURI, an AI attendance system, and is accompanied by a comprehensive training pipeline, pretrained weights, and ONNX export capabilities. The model classifies faces into two categories: Real or Spoof, with an impressive operational precision exceeding 99%. The MiniFAS architecture offers significant advantages over previous iterations, including reduced model size and faster inference while maintaining high accuracy metrics validated against the CelebA Spoof benchmark. Significantly, the MiniFAS leverages advanced techniques such as the Fourier Transform auxiliary loss to enhance its ability to learn discerning texture patterns that differentiate between genuine skin and artificial images. The model has been optimized for various deployment scenarios, producing quantized versions that minimize storage requirements to just 600 KB without sacrificing accuracy. This development is particularly impactful for the AI/ML community, as it can facilitate greater adoption of secure biometric solutions across various platforms and applications, ultimately enhancing trust and security in automated systems.