Improving Monocular Facial Presentation-Attack-Detection Robustness with Synthetic Noise Augmentations.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Face presentation-attack-detection (PAD) is crucial for secure authentication, protecting against spoofing methods like photos and videos.
• Current high-performance PAD systems often rely on expensive 3D imaging, driving research into cost-effective monocular (single-camera) approaches.
• Monocular PAD methods excel in controlled settings but struggle with real-world imaging variations such as sensor noise and dynamic lighting.

Purpose of the Study:

• To develop a synthetic augmentation strategy to improve the robustness of monocular face PAD against real-world noise.
• To create a practical and cost-effective solution for enhancing PAD performance without extensive data collection under all conditions.

Main Methods:

• Introduced a physics-informed noise augmentation toolbox with twelve generators simulating sensor and lighting effects.
• Evaluated the generated features against popular augmentation methods on noisy test datasets.
• Conducted an ablation study to quantify the impact of synthetic augmentations versus real data.

Main Results:

• The proposed synthetic noise augmentations generated more robust PAD features compared to existing methods in noisy evaluations.
• The augmentation toolbox also improved accuracy on clean test data, indicating better discrimination between spoof and imaging artifacts.
• Achieved superior test accuracy with only 30% of participants requiring full imaging, demonstrating significant data collection simplification.

Conclusions:

• Physics-informed synthetic noise augmentations offer a pragmatic approach to enhance monocular face PAD robustness.
• This method effectively addresses real-world noise challenges, leading to improved security and reduced data acquisition costs.
• Synthetic augmentations are a valuable tool for developing more resilient and accessible biometric authentication systems.