Deep Learning for Face Anti-Spoofing: A Survey.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face anti-spoofing (FAS) is crucial for securing face recognition systems against presentation attacks (PAs).
• Traditional handcrafted feature methods are insufficient for novel and realistic PAs.
• Deep learning (DL) based FAS has become dominant due to large datasets and superior performance.

Purpose of the Study:

• To provide the first comprehensive review of recent deep learning-based face anti-spoofing methods.
• To stimulate future research by highlighting novel approaches and open challenges in FAS.
• To bridge the gap between outdated reviews focusing on handcrafted features and current DL advancements.

Main Methods:

• Reviewing DL-based FAS methods, including those using pixel-wise supervision (e.g., pseudo depth maps) beyond binary labels.
• Analyzing methods for domain generalization and open-set face anti-spoofing, moving beyond traditional intra-dataset evaluations.
• Summarizing DL applications for multi-modal sensors (depth, infrared) and specialized sensors (light field, flash) alongside RGB cameras.

Main Results:

• Deep learning methods significantly outperform traditional approaches in face anti-spoofing.
• Novel supervision techniques and domain generalization strategies show promise for robust FAS.
• Multi-modal and specialized sensors offer new avenues for enhancing face anti-spoofing capabilities.

Conclusions:

• Deep learning has revolutionized face anti-spoofing, offering more robust solutions.
• Future research should focus on advanced supervision, domain generalization, and diverse sensor modalities.
• Addressing current open issues is key to advancing the field of face anti-spoofing.