Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Supervised Contrastive Learning for Snapshot Spectral Imaging Face Anti-Spoofing

Chuanbiao Song, Yan Hong, Jun Lan, Huijia Zhu, Weiqiang Wang, Jianfu Zhang

TL;DR

This work tackles face anti-spoofing under snapshot spectral imaging by introducing a re-balanced supervised contrastive learning framework that combines data resampling, intra-class mixup, and a Real-face Oriented Reweighting strategy to reduce identity bias. The method processes multimodal inputs (RGB $\oplus$ 30 spectral channels) with a spectral-weighted, multi-attention backbone and optimizes a dual objective: $\mathcal{L} = \mathcal{L}_{\text{c}} + \lambda_{scl} \cdot \mathcal{L}_{\text{scl}}$ (with $\lambda_{scl}=10$), including Focal Loss and a Supervised Contrastive Loss aided by Cross Batch Memory, using $\tau=0.07$ for temperature. On the HySpeFAS dataset, the approach achieves a record ACER of $0.0000\%$, with APCER and BPCER also at $0\%$, ranking 1st in the CVPR 2024 challenge. The findings demonstrate that aligned contrastive learning, proper resampling, and bias-aware weighting can substantially boost SSI-based FAS performance with practical implications for secure biometric systems.

Abstract

This study reveals a cutting-edge re-balanced contrastive learning strategy aimed at strengthening face anti-spoofing capabilities within facial recognition systems, with a focus on countering the challenges posed by printed photos, and highly realistic silicone or latex masks. Leveraging the HySpeFAS dataset, which benefits from Snapshot Spectral Imaging technology to provide hyperspectral images, our approach harmonizes class-level contrastive learning with data resampling and an innovative real-face oriented reweighting technique. This method effectively mitigates dataset imbalances and reduces identity-related biases. Notably, our strategy achieved an unprecedented 0.0000\% Average Classification Error Rate (ACER) on the HySpeFAS dataset, ranking first at the Chalearn Snapshot Spectral Imaging Face Anti-spoofing Challenge on CVPR 2024.

Supervised Contrastive Learning for Snapshot Spectral Imaging Face Anti-Spoofing

TL;DR

This work tackles face anti-spoofing under snapshot spectral imaging by introducing a re-balanced supervised contrastive learning framework that combines data resampling, intra-class mixup, and a Real-face Oriented Reweighting strategy to reduce identity bias. The method processes multimodal inputs (RGB 30 spectral channels) with a spectral-weighted, multi-attention backbone and optimizes a dual objective: (with ), including Focal Loss and a Supervised Contrastive Loss aided by Cross Batch Memory, using for temperature. On the HySpeFAS dataset, the approach achieves a record ACER of , with APCER and BPCER also at , ranking 1st in the CVPR 2024 challenge. The findings demonstrate that aligned contrastive learning, proper resampling, and bias-aware weighting can substantially boost SSI-based FAS performance with practical implications for secure biometric systems.

Abstract

This study reveals a cutting-edge re-balanced contrastive learning strategy aimed at strengthening face anti-spoofing capabilities within facial recognition systems, with a focus on countering the challenges posed by printed photos, and highly realistic silicone or latex masks. Leveraging the HySpeFAS dataset, which benefits from Snapshot Spectral Imaging technology to provide hyperspectral images, our approach harmonizes class-level contrastive learning with data resampling and an innovative real-face oriented reweighting technique. This method effectively mitigates dataset imbalances and reduces identity-related biases. Notably, our strategy achieved an unprecedented 0.0000\% Average Classification Error Rate (ACER) on the HySpeFAS dataset, ranking first at the Chalearn Snapshot Spectral Imaging Face Anti-spoofing Challenge on CVPR 2024.
Paper Structure (18 sections, 7 equations, 3 figures, 2 tables)

This paper contains 18 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. HySpeFAS dataset
  3. Methodology
  4. Data Preprocessing
  5. Class Balancing
  6. Data Augmentations
  7. Framework
  8. Training Strategies
  9. Intra-class Mixup
  10. Real-face Oriented Reweighting
  11. Objective Functions
  12. Focal Loss.
  13. Supervised Contrastive Loss.
  14. Experiments
  15. Evaluation Metrics
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Analysis of examples from HySpeFAS dataset. From top to bottom: real images and fake images. From left to right: examples in the first box show the same ID face across real and fake images; the second one indicates the various alternations of real and fake images; the third one visualizes different appearances of images from the same ID; the last one shows different orientations of real images and fake images.
  • Figure 2: The framework of our proposed method.
  • Figure 3: Visualization of augmented examples with three different type of random mask.