Facial Liveness Detection
Lightweight face anti-spoofing system for government-scale identity verification
Designed and shipped a 600KB on-device liveness model reaching 97.8% accuracy with 5–10ms CPU inference on mobile devices. The system detects 89.5% of presentation attacks including printed photos, screen replays, and masks, while maintaining zero false rejects at the selected production threshold. Deployed in production on Android (ONNX Runtime) and iOS (CoreML) for government identity verification flows at Nashid.
The hard part
Compressing the model 3x (1.82MB → 600KB) without losing accuracy required careful ONNX quantization with representative calibration data. Threshold tuning to find the optimal operating point that maximizes fraud detection while maintaining zero false rejections.
What I did
End-to-end ownership: Owned the system end-to-end: designed and trained the liveness model, built the data and training pipeline, optimized the model for on-device performance, and led it from experimentation to real production use. Worked closely with mobile and product teams to integrate the model into live verification flows and set up monitoring to track performance and reliability in the field.