Convolutional Neural Network Approach for Multispectral Facial Presentation Attack Detection in Automated Border...

Area of Science:

• Computer Science
• Biometrics
• Security Engineering

Background:

• Automated border control systems are crucial for national security, facing risks from unauthorized passenger crossings.
• Facial biometrics are widely used but vulnerable to presentation attacks (e.g., printed or displayed images).
• Existing research often relies on laboratory conditions, not reflecting real-world physiological variations.

Purpose of the Study:

• To develop and evaluate a multispectral facial biometrics system for robust presentation attack detection (PAD).
• To create a novel, in situ acquired database using visible (VIS), near-infrared (NIR), and thermal imaging.
• To assess the effectiveness of different sensor fusion techniques and classifiers under realistic conditions.

Main Methods:

• Acquired a new multispectral facial image database in situ, capturing VIS, NIR, and thermal data.
• Investigated presentation attacks including printed, masked, and displayed images.
• Employed convolutional neural networks (CNNs) for feature extraction and evaluated five classifiers (e.g., KNN, SVM) for PAD.
• Implemented both classifier-level and feature-level sensor fusion strategies.

Main Results:

• Thermal sensors demonstrated superior performance individually compared to other sensors.
• Combining data from all three sensors (VIS, NIR, thermal) significantly improved PAD accuracy, irrespective of fusion method.
• Classifiers such as K-Nearest Neighbors (KNN) and Support Vector Machines (SVM) achieved high performance with low computational cost.
• The in situ database acquisition addressed limitations of laboratory-controlled environments.

Conclusions:

• Multispectral facial biometrics, particularly with thermal imaging and sensor fusion, offers enhanced security for automated border control.
• Real-world physiological variations do not impede the effectiveness of the proposed multispectral PAD system.
• The developed system provides a practical and computationally efficient solution for detecting sophisticated presentation attacks.