Deep Hybrid Multimodal Biometric Recognition System Based on Features-Level Deep Fusion of Five Biometric Traits.

Area of Science:

• Computer Science
• Biometrics
• Information Security

Background:

• Increasing global demand for robust information security and regulatory compliance.
• Limitations of unimodal biometric systems in accuracy and security.
• Hybrid multimodal biometric systems offer improved recognition by fusing multiple traits.

Purpose of the Study:

• To propose and compare three novel feature-level deep fusion strategies for hybrid multimodal biometric systems.
• To enhance recognition accuracy, security assurance, and robustness against spoofing.
• To address the limitations of unimodal systems through effective trait fusion.

Main Methods:

• Feature-level deep fusion of five biometric traits (face, irises, fingerprints) from three sources.
• Mapping feature vectors to Reproducing Kernel Hilbert Space (RKHS) for nonlinear-to-linear conversion.
• Utilizing dimensionality reduction (KPCA, KLDA) and quaternion-based algorithms (KQPCA) in RKHS.
• Employing deep learning with fully connected layers for feature space fusion.

Main Results:

• Experimental validation on 6 databases demonstrated the effectiveness of the proposed hybrid multimodal biometric system.
• The deep fusion approach resulted in a secure and robust multimodal template.
• Achieved 100% accuracy in the hybrid multimodal biometric system, surpassing unimodal and other multimodal systems.

Conclusions:

• Deep fusion of feature spaces in hybrid multimodal biometric systems significantly enhances recognition accuracy and robustness.
• The proposed methods provide a secure and efficient solution for information security challenges.
• The low dimensionality of the fused vector contributes to increased system performance and spoof attack resistance.