Quality-Aware Secure Deep Fusion Model for Multimodal Biometric Authentication

Biometric authentication is a reliable alternative to password- and token-based access control; however, unimodal biometric systems suffer from noisy samples, intra-class variation, non-universality, spoofing, and increased false acceptance and false rejection rates. This paper presents a Quality-Aware Secure Deep Fusion Model for multimodal biometric authentication using fingerprint and palmprint traits. The model strengthens an existing DNN-MLP multimodal biometric baseline by adding sample-quality assessment, liveness detection, enhanced preprocessing, CNN/Vision Transformer feature extraction, cancellable encrypted template protection, adaptive score-level fusion, and dynamic threshold selection. Fingerprint and palmprint samples are first screened for blur, noise, ridge clarity, illumination, region-of-interest completeness, and texture quality. Live samples are enhanced and transformed into discriminative biometric embeddings, while protected templates are stored using cancellable and encrypted mechanisms. During authentication, modality-specific scores are normalized and fused according to quality, liveness confidence, and modality reliability. The comparative evaluation shows improved accuracy from 97.60% to 99.20%, reduced FAR from 1.30% to 0.36%, and improved FRR, TAR, precision, recall, F1-score, EER, and ROC-AUC. The proposed framework provides a more accurate, robust, and secure solution for real-time multimodal biometric authentication.