Optimizing Deep CNN Architectures for Face Liveness Detection.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Facial recognition systems are vulnerable to spoofing attacks using photographs.
• Face liveness detection is crucial for secure biometric authentication.
• Existing methods require enhancement to reliably distinguish real faces from fakes.

Purpose of the Study:

• To develop advanced deep learning architectures for robust face liveness detection.
• To improve upon existing texture analysis and convolutional neural network (CNN) approaches.
• To evaluate and compare different deep architectures for optimal performance.

Main Methods:

• Utilized nonlinear diffusion to enhance image texture and edges.
• Employed deep convolutional neural networks (CNNs) for feature extraction and classification.
• Evaluated deep CNN, residual network, and Inception Network v4 architectures on the NUAA dataset.

Main Results:

• Achieved 100% accuracy in face liveness detection on the NUAA Photograph Impostor dataset with an enhanced architecture.
• Inception Network v4 demonstrated optimal performance with nonlinear anisotropic diffused images.
• The proposed method surpassed current state-of-the-art face liveness detection techniques.

Conclusions:

• Deep learning architectures combined with image enhancement offer a highly effective solution for face liveness detection.
• Inception Network v4 is particularly well-suited for this task, providing superior accuracy.
• The developed approach significantly enhances the security of facial recognition systems against spoofing.