Skip to content

Face Anti-Spoofing Feature Extraction

Definition

This article covers the specific features and cues that liveness detection models use to distinguish real faces from spoofs — texture, frequency, depth, reflection, and temporal features.

Feature Categories

Feature Type What It Captures Detection Target
Texture Surface patterns, print dots, moiré, skin pores Print, screen replay
Frequency Spectral characteristics, high-frequency details GAN artifacts, print patterns
Depth 3D structure of face 2D attacks (flat)
Reflection Specular reflection patterns Material differences (skin vs paper/screen)
Color Color distribution, chrominance Screen color gamut, paper white point
Temporal Motion patterns, flickering, blinking Video replay, deepfake inconsistency
Contextual Background, edges, device frame Screen bezel, paper edges

Texture Features

Technique How It Works
LBP (Local Binary Patterns) Encode local texture patterns — paper/screen have different micro-texture than skin
Central Difference Convolution Capture gradient information that regular convolution misses — key innovation of CDCN
Gabor filters Multi-scale, multi-orientation texture analysis
Deep CNN features Learned texture representations from convolutional layers

Frequency Features

Technique How It Works
FFT/DFT analysis Moiré patterns create distinctive frequency peaks
Wavelet decomposition Multi-resolution frequency analysis
High-frequency analysis Real faces have fine detail; prints/screens lose high-frequency information

Depth Features

Technique How It Works
Monocular depth estimation Predict depth map from single RGB image — real face is 3D, spoof is flat
Structured light Project pattern, measure distortion for depth (iPhone FaceID)
Stereo depth Two cameras compute depth via parallax

Key Takeaways

Summary

  • Multiple feature types work together — no single feature catches all attacks
  • Texture + depth is the most effective combination for RGB-only systems
  • Central Difference Convolution is the key architectural innovation for fine-grained texture
  • Frequency analysis is particularly effective against print attacks (moiré detection)
  • Temporal features add significant value when video is available