Biometric authentication is now crucial for secure digital access. However, unimodal methods still face issues like spoofing, background noise, and high cloud delays. Earlier multimodal models offered improved reliability but struggled with high computational demands and moderate accuracy levels (around 85-90%). This paper presents a new multimodal biometric authentication framework. It uses Deep Image Prior (DIP) for preprocessing, lightweight CNNs for feature extraction, and edge computing for real-time deployment. To boost performance, the model includes a Convolutional Block Attention Module (CBAM) and a Capsule Network layer. These components enhance the learning of unique features across fingerprint, iris, and facial types. The extracted features are combined using Fisher Vector with Gaussian Mixture Models (GMM) and classified through a quantized ResNet-101 backbone. Tests on the CASIA multimodal dataset, which includes over 10,000 samples, show that the proposed model reaches 96% accuracy. It surpasses current unimodal and multimodal systems and reduces latency by 40% on Raspberry Pi 4 and Jetson Nano platforms. The use of attentionguided and capsule layers sets this method apart from earlier models, providing a scalable solution that resists spoofing for banking, healthcare, defense, and IoT applications Index Terms—Biometric Authentication, Deep Image Prior, Multimodal Fusion, Edge Computing, Attention Mechanism, Capsule Networ
R et al. (Thu,) studied this question.