Adaptive management of multimodal biometrics—A deep learning and metaheuristic approach

Abstract This paper introduces the framework for adaptive rank-level biometric fusion: a new approach towards personal authentication. In this work, a novel attempt has been made to identify the optimal design parameters and framework of a multibiometric system, where the chosen biometric traits are subjected to rank-level fusion. The optimal fusion parameters depend upon the security level demanded by a particular biometric application. The proposed framework makes use of a metaheuristic approach towards adaptive fusion in the pursuit of achieving optimal fusion results at varying levels of security. Rank-level fusion rules have been employed to provide optimum performance by making use of Ant Colony Optimization technique. The novelty of the reported work also lies in the fact that the proposed design engages three biometric traits simultaneously for the first time in the domain of adaptive fusion, so as to test the efficacy of the system in selecting the optimal set of biometric traits from a given set. Literature reveals the unique biometric characteristics of the fingernail plate, which have been exploited in this work for the rigorous experimentation conducted. Index, middle and ring fingernail plates have been taken into consideration, and deep learning feature-sets of the three nail plates have been extracted using three customized pre-trained models, AlexNet, ResNet-18 and DenseNet-201. The adaptive multimodal performance of the three nail plates has also been checked using the already existing methods of adaptive fusion designed for addressing fusion at the score-level and decision-level. Exhaustive experiments have been conducted on the MATLAB R2019a platform using the Deep Learning Toolbox. When the cost of false acceptance is 1.9, experimental results obtained from the proposed framework give values of the average of the minimum weighted error rate as low as 0.0115, 0.0097 and 0.0101 for the AlexNet, ResNet-18 and DenseNet-201 based experiments respectively. Results demonstrate that the proposed system is capable of computing the optimal parameters for rank-level fusion for varying security levels, thus contributing towards optimal performance accuracy.