Adaptive Multimodal Biometric Recognition Framework Integrating Iris and Fingerprint Modalities for Robust and Interpretable Authentication
S. Ramesh¹, V. Krishnaveni²

¹S. Ramesh, Research Scholar, Department of Electronics and Communication Engineering, PSG College of Technology, Coimbatore (Tamil Nadu), India.

²Dr. V. Krishnaveni, Professor & Head, Department of Electronics and Communication Engineering, PSG College of Technology, Coimbatore (Tamil Nadu), India.    

Manuscript received on 23 September 2025 | First Revised Manuscript received on 05 October 2025 | Second Revised Manuscript received on 18 October 2025 | Manuscript Accepted on 15 November 2025 | Manuscript published on 30 November 2025 | PP: 1-12 | Volume-5 Issue-4, November 2025 | Retrieval Number: 100.1/ijsp.D101805041125 | DOI: 10.54105/ijsp.D1018.05041125

Open Access | Editorial and Publishing Policies | Cite | Zenodo | OJS |  Indexing and Abstracting
© The Authors. Published by Lattice Science Publication (LSP). This is an open-access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Biometric recognition is a vital technology for secure identification. However, unimodal systems often face several drawbacks, including reduced reliability in challenging demographic groups such as children, environmental constraints, and susceptibility to spoofing. To address these problems, this work presents a scalable multimodal biometric architecture that integrates fingerprint and iris modalities to enhance identification accuracy, resilience, and interpretability. In the proposed architecture, unique iris patterns are captured using a Multi-Layer Perceptron Mixer (MLP-Mixer), and concise yet discriminative fingerprint attachments are generated using a Variational Autoencoder (VAE). Using a Triplet Network to improve the distinction between real and fake samples, the matching performance is further strengthened. Appropriate blending is achievd through feacture-level fusion using a Cross-Attention Transformer, which dynamically aligns complementary iris and fingerprint embeddings. Crucially, by highlighting each feature’s contribution to the final decision, Integrated Gradient (IG) are used to guarantee integritu and openness.. The efficacy of this technique is demonstrated by experiments using benchmark iris and fingerprint datasets, which achieved an overall identification accuracy of 97.8%, with Equal Error Rates (EER) of 0.6% for iris and 0.7% for fingerprints. The robustness of the suggested paradigm is demonstrated by comparison against unimodal and multimodal baselines, especially in situations involving noisy data and early-age biometric identification. These results highlight the system’s usefulness for applications in safeguarding children, scalable authentication, and encrypted control of entry. In the final analysis, this study not only creates a new path for interpretable multisensory fusion but also creates the foundation for extending biometric solutions to larger populations and operational settings.

Keywords: CASIA Iris Datasets, NIST Special Databases, MLPMixer, Variational Autoencoders, Triplet Network, CrossAttention Transformers, Integrated Gradients.
Scope of the Article: Machine Vision