5th Face Anti-spoofing Workshop @ CVPR2024

Event: 5th Face Anti-spoofing Workshop @ CVPR2024 · Duration: 52 min · ▶ Watch on YouTube

Abstract

This workshop explores the critical area of face anti-spoofing, addressing vulnerabilities in remote biometric authentication systems. It delves into various attack vectors, including presentation attacks and digital injection attacks like deepfakes and adversarial samples, highlighting the limitations of current passive anti-spoofing methods. The presentations introduce novel approaches such as proactive anti-spoofing systems that leverage sensor features and two-stage training strategies with micro-disturbance techniques to enhance robustness against unseen attack types. Additionally, the workshop showcases advanced imaging techniques like Snapshot Spectral Imaging and the HySpeFAS dataset, demonstrating how spectral-spatial feature fusion and data augmentation can overcome data challenges in face anti-spoofing.

Speakers

  • Xiang Xu — AWS AI Labs
  • Fengjun Guo — INTSIG Information Co. Ltd.
  • Hui Li — China Telecom Artificial Intelligence Technology Co. Ltd.

Talks (3)

  • 00:00:00 — Xiang Xu: Principles of Designing Remote Face Anti-Spoofing Systems
    • This talk discusses vulnerabilities in remote biometric authentication systems, categorizing them into presentation attacks and digital injection attacks (deepfakes, adversarial samples, digital replay), and proposes a proactive anti-spoofing approach leveraging sensor features to increase attack costs and improve robustness.
  • 02:25:00 — Fengjun Guo: Unified Face Attack Detection with Micro Disturbance and a Two-Stage Training Strategy
    • This talk introduces a two-stage training strategy for face attack detection, utilizing a pre-training stage with masked input images and a fine-tuning stage with a Swin-Transformer, along with a ‘Micro Disturbance’ method to expand data diversity and improve cross-domain generalization for unseen attack types.
  • 03:40:00 — Hui Li: Snapshot Spectral Imaging for Face Anti-Spoofing: Addressing Data Challenges with Advanced Processing and Training
    • This talk explores Snapshot Spectral Imaging (SSI) for face anti-spoofing, introducing the HySpeFAS dataset and proposing a spectral-spatial feature fusion network with spectral-spatial data augmentation to address data challenges and improve performance in capturing high-dimensional spectral information.

Key Takeaways

  • Passive face anti-spoofing is not reliable, especially with single-image methods, and security depends on the weakest modules in the system.
  • Proactive anti-spoofing approaches, leveraging sensor features and random signals, can significantly increase the cost for attackers and improve robustness against various attack types, including digital replay attacks.
  • A two-stage training strategy (pre-training with reconstruction-based MIM and fine-tuning with Swin-Transformer) combined with a Micro Disturbance method can effectively enhance a model’s ability to detect unseen attack types and improve generalization.
  • Snapshot Spectral Imaging (SSI) offers high-dimensional spectral information, and challenges like limited data can be addressed through spectral-spatial feature fusion and data augmentation, as demonstrated with the HySpeFAS dataset.

Methods / Models / Datasets Mentioned

  • PGD
  • SimBA
  • CDCN
  • MDFAS
  • SAFAS
  • CADM
  • Resnet-50
  • Swin-Transformer
  • MIM (Masked Image Modeling)
  • SimMIM
  • Grad-CAM
  • HySpeFAS dataset
  • Spectral-Spatial Feature Fusion Network
  • Spectral-Spatial Data Augmentation
  • t-SNE

Topics

Face Anti-Spoofing · Remote Biometric Authentication · Presentation Attacks · Digital Injection Attacks · Deepfakes · Adversarial Samples · Digital Replay Attacks · Proactive Anti-Spoofing · Two-Stage Training · Micro Disturbance · Snapshot Spectral Imaging (SSI) · Spectral-Spatial Feature Fusion · Data Augmentation · Cross-Domain Generalization

Notes

Open for commentary — connections to other work, critiques, follow-up reading.