Advanced Deep Learning Architecture for Multimodal Biometric Authentication Using Feature-Level Fusion and Dimensionality Reduction

Multimodal biometric authentication systems represent a critical advancement in secure identity verification for critical infrastructure and sensitive applications. This research presents a novel deep learning architecture combining Histogram of Oriented Gradients (HOG) features with deep learning models (VGG16 for fingerprints, FaceNet for faces) integrated with Principal Component Analysis (PCA) dimensionality reduction and Fully Connected Neural Network (FCNN) classification. Our experimental results demonstrate 98.3% accuracy for fingerprint recognition and 97.6% accuracy for face recognition when using the proposed FCNN classifier. The feature-level fusion approach addresses the computational challenges inherent in high dimensional biometric data while maintaining superior classification performance compared to traditional machine learning methods including Support Vector Machines (SVM), Random Forests, and Convolutional Neural Networks (CNNs). Comparative analysis reveals that our FCNN-based approach outperforms gradient boosting (96.8% fingerprint, 95.3% face) and conventional CNN methods (96.2% fingerprint, 95.4% face). The system incorporates advanced regularization techniques including dropout (rate: 0.5) and L2 regularization to prevent overfitting. Through 25 epochs of training, the system achieves convergence with training accuracy reaching 98.3% and validation accuracy stabilizing at 98.1%, demonstrating robust generalization capabilities. Our contributions include: (1) a novel feature fusion architecture combining handcrafted and learned representations, (2) comprehensive PCA analysis preserving 95.6% variance with only 110 components, and (3) systematic evaluation of multiple classification paradigms demonstrating FCNN superiority. This work advances the state-of-the-art in multimodal biometric authentication, offering practical implications for cloud-based access control systems and high-security applications.

Keywords: Biometric authentication, deep learning, feature fusion, dimensionality reduction, neural networks, fingerprint recognition, face recognition, PCA, FCNN