Skip to content

On-Device Biometric Processing

Definition

On-device biometric processing runs face detection, liveness detection, and/or recognition models directly on the user's phone, rather than sending images to a remote server. This provides faster feedback, better privacy, offline capability, and lower infrastructure costs.

On-Device ML Frameworks

Framework Platform Key Feature
CoreML iOS Apple's native ML inference, Neural Engine support
TensorFlow Lite Android/iOS Cross-platform, GPU delegate, NNAPI
ONNX Runtime Mobile Android/iOS Cross-framework model compatibility
MediaPipe Android/iOS/Web Google's real-time ML pipeline
MNN Android/iOS Alibaba's mobile inference engine
NCNN Android Tencent's efficient mobile inference
PyTorch Mobile Android/iOS PyTorch's mobile deployment

Model Optimization for Mobile

Technique What It Does Typical Speedup
Quantization (INT8) Reduce weights from FP32 to INT8 2-4x faster, 4x smaller
Pruning Remove unimportant weights 1.5-3x faster
Knowledge distillation Train small model to mimic large model Smaller model with near-large accuracy
Architecture search Design mobile-specific architectures MobileFaceNet, GhostNet
Operator fusion Combine multiple operations into one 1.2-1.5x faster

Typical On-Device Performance

Task Model Size Latency (Mobile)
Face detection SCRFD-500M 2.5MB 5-15ms
Face detection BlazeFace 0.2MB 1-3ms
Face recognition MobileFaceNet 4MB 15-30ms
Face liveness MobileNetV3-Small 6MB 10-25ms
Full pipeline Detection + Liveness + Recognition ~15MB 30-80ms 
graph TD
    A[Camera] --> B[On-Device]
    B --> B1[Face detection]
    B --> B2[Quality check]
    B --> B3[Quick liveness]
    B --> B4[Guide user in real-time]

    B1 & B2 & B3 --> C{On-device pass?}
    C -->|No| D[Immediate retry guidance]
    C -->|Yes| E[Send to Server]

    E --> F[Server Processing]
    F --> F1[Deep liveness analysis]
    F --> F2[Face recognition + matching]
    F --> F3[Document processing]
    F --> F4[Database verification]

    F1 & F2 & F3 & F4 --> G[Final Decision]

    style B fill:#2E7D32,color:#fff
    style F fill:#4051B5,color:#fff

Key Takeaways

Summary

  • On-device processing provides instant feedback, better privacy, and offline capability
  • ONNX Runtime and TFLite are the most versatile cross-platform options
  • INT8 quantization gives 2-4x speedup with minimal accuracy loss
  • MobileFaceNet (4MB, 15-30ms) is the standard for on-device face recognition
  • Hybrid architecture is the practical approach — on-device for speed + server for depth
  • Full on-device pipeline runs in 30-80ms on modern phones