Enhanced Deep Learning Architectures for Face Liveness Detection for Static and Video Sequences.

Area of Science:

• Computer Vision and Machine Learning
• Biometric Security Systems

Background:

• Face liveness detection is crucial for preventing identity fraud in face recognition systems.
• Existing high-accuracy methods are too slow for real-time applications due to multi-step processing.

Purpose of the Study:

• To develop and evaluate end-to-end, real-time face liveness detection solutions.
• To integrate image diffusion techniques with deep learning for improved spoofing attack detection.

Main Methods:

• Applied nonlinear anisotropic diffusion to enhance image features and preserve boundaries.
• Utilized Specialized Convolutional Neural Network (SCNN) and Inception v4 for static image liveness classification.
• Developed a novel deep architecture combining diffusion with Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) for video sequence analysis.

Main Results:

• Achieved high accuracies: 96.03% (SCNN) and 94.77% (Inception v4) on Replay-Attack dataset.
• Achieved high accuracies: 96.21% (SCNN) and 95.53% (Inception v4) on Replay-Mobile dataset.
• Video analysis model yielded 98.71% accuracy on Replay-Attack and 95.41% on Replay-Mobile.

Conclusions:

• The proposed end-to-end diffusion-based deep learning models enable real-time face liveness detection.
• The integrated approach significantly improves accuracy and efficiency in combating face spoofing attacks.
• The novel CNN-LSTM architecture combined with diffusion offers a robust solution for video-based liveness detection.