Security-aware user authentication based on multimodal biometric data using dilated adaptive RNN with optimal weighted feature fusion

This work plans to develop a biometric authentication model by the combination of multi-modal inputs like voice, fingerprint, and iris to provide high security. At first, the spectrogram images, the collected fingerprint, and the collected iris input were given to a Multi-scale Residual Attention Network (RAN) with Atrous Spatial Pyramid Pooling (ASPP) to extract the best values. These three features are then fed to optimal weighted feature fusion, where weight optimization from the features is done via the Enhanced Lichtenberg Algorithm (ELA). These features are fed into the decision-making stage, where the Dilated Adaptive Recurrent Neural Network is utilized to identify the individuals, where the parameters are optimized from RNN using ELA to improve the recognition performance. The simulation findings achieved from the developed multimodal authentication systems are validated using diverse algorithms over several efficacy metrics like accuracy, precision, sensitivity, F1-score, etc. From the result analysis, the ELA-DARNN-based user authentication system showed a higher accuracy of 96.01, and other models such as 90% than SVM, CNN, CNN-AlexNet, and Dil-ARNN given the accuracy to be 87.94, 89.88, 93.25, and 91.94. Therefore, the outcomes explored that the offered approach has attained elevated results and also effectively supports to reduction of data theft.