Multimodal Biometric Authentication using Siamese Network-based Metric Learning

Biometric authentication systems have also become one of the pillars of the contemporary security systems because of their capability of offering reliable and user-friendly identity verification. The unimodal biometric systems of the past, which use a single characteristic like fingerprint or even face, have been known to have problems with performance reduced to noisy conditions of acquisition, spoofing attacks, and intra-class variation. Multimodal biometric authentication will overcome these shortcomings by combining more than one biometric characteristic hence enhancing strength and accuracy. However, in recent years, metrics learning algorithms based on deep learning, especially Siamese networks, have shown a high level of ability to learn discriminative embeddings in verification tasks. The paper introduces a framework of multimodal biometric authentication using a Siamese network, which is aimed at learning joint and modality-invariant feature representations of heterogeneous biometric modalities. The proposed architecture builds on parallel modality-specific encoders and then a Siamese architecture to learn similarities using different conditions of acquisition. The architecture of the system, training plan as well as the fusion mechanism are addressed. An extensive discussion of the related literature, issues, and the scope of future studies is also done. The proposed structure will enhance the accuracy of verification, scalability and scalability of multimodal biometric systems in real life.