A Novel Token-wise Asymmetric Contrastive Learning For Robust Face Presentation Attack Detection

Abstract

Face Recognition Systems Have Become A Fundamental Component Of Modern Biometric Authentication. However, These Systems Remain Vulnerable To Presentation Attacks Such As Printed Photographs, Replay Videos, Masks, And AI-generated Deepfakes. Existing Face Anti-spoofing (FAS) Methods Often Exhibit Limited Generalization When Exposed To Unseen Attack Types And Cross-domain Variations. This Paper Presents A Robust FAS Framework That Combines Token-level Feature Learning, Contrastive Representation Learning, And Angular Margin Optimization To Improve Liveness Detection Performance. The Proposed Framework Learns Discriminative Feature Representations By Encouraging Compact Live-face Embeddings And Enhanced Separation From Spoof-face Embeddings. Localized Facial Analysis Enables The Extraction Of Fine-grained Liveness Cues, Including Texture Inconsistencies, Illumination Artifacts, And Reflectance Variations. Experimental Evaluation Demonstrates An Accuracy Of 90.91%, An ACER Of 9.9%, And A ROC-AUC Of 0.9903, Indicating Strong Discriminative Capability. The Proposed System Provides An Effective And Practical Solution For Enhancing Biometric Security Against Presentation Attacks.

Keywords