Skip to content

Liveness Model Architectures

Definition

This article covers the deep learning architectures specifically designed for face anti-spoofing / liveness detection, including CNN-based, transformer-based, and hybrid approaches with auxiliary supervision.

Architecture Categories

graph TD
    A[Liveness Architectures] --> B[Binary Classification<br/>Real vs Spoof]
    A --> C[Pixel-wise Supervision<br/>Depth/Reflection maps]
    A --> D[Multi-Task Learning<br/>Spoof type + binary]
    A --> E[Domain-Generalized<br/>Cross-dataset robust]
    A --> F[Foundation Model<br/>Pre-trained + fine-tuned]

    style E fill:#4051B5,color:#fff
    style F fill:#6A1B9A,color:#fff

Key Architectures

CDCN (Central Difference Convolution Network)

Aspect Details
Paper Searching Central Difference Convolutional Networks for FAS (CVPR 2020)
Key innovation Central Difference Convolution โ€” captures fine-grained gradient patterns that regular convolution misses
Supervision Pixel-wise depth map estimation
Why it works Real faces have 3D depth structure; spoofs are flat โ€” CDC is sensitive to these subtle gradient differences
Performance Strong cross-dataset results, small model size (~2M params)

ViT for Liveness

Aspect Details
Approach Vision Transformer with self-attention captures global patterns
Benefit Attends to both local (texture) and global (layout, context) spoof cues
Models ViT-Small (22M), ViT-Tiny (5.7M)
Patch-based Each image patch is a token โ€” model learns which patches contain spoof evidence

Auxiliary Supervision Tasks

Auxiliary Task What the Model Predicts Why It Helps
Depth map Per-pixel depth of face Real face = 3D structure, spoof = flat
Reflection map Specular reflection patterns Paper/screen reflect differently than skin
Binary mask Which pixels are face vs spoof medium Fine-grained spatial understanding
Domain label Which dataset/domain the sample is from Adversarial: forces domain-invariant features 
graph TD
    A[Face Image 256ร—256] --> B[Shared Backbone<br/>ResNet-18 / CDCN]
    B --> C[Depth Head<br/>Predicts face depth map]
    B --> D[Classification Head<br/>Real vs Spoof]
    B --> E[Reflection Head<br/>Predicts reflection map]

    C --> F[Depth Loss<br/>MSE with ground truth]
    D --> G[BCE Loss<br/>Binary cross-entropy]
    E --> H[Reflection Loss]

    F & G & H --> I[Total Loss = ฮป1ยทL_depth + ฮป2ยทL_cls + ฮป3ยทL_ref]

    style B fill:#4051B5,color:#fff

Model Comparison

Model Params OULU-NPU (ACER%) Cross-Dataset Mobile-Ready
CDCN 2M 1.0% Good โœ…
ResNet-18 + Binary 11M 3-5% Fair โœ…
ViT-Small 22M 1-2% Good โš ๏ธ
SSDG (ResNet-18) 11M 2-4% Very Good โœ…
FLIP-MCL Large 0.5-1% Excellent โŒ

Key Takeaways

Summary

  • CDCN introduced Central Difference Convolution โ€” captures fine gradient patterns missed by regular CNN
  • Auxiliary supervision (depth maps, reflection maps) significantly outperforms pure binary classification
  • ViT brings global attention to liveness โ€” can attend to context cues across the full image
  • Multi-task learning with depth + binary is the most common production approach
  • Model size matters for mobile: CDCN (2M) and ResNet-18 (11M) are practical; ViT-Base (86M) is not