Multimodal biometric fusion using evolutionary techniques

Abstract

The work of this research focuses on fusing multiple biometric modalities at the score level using different combination rules. The research puts an emphasis on employing optimization techniques in order to achieve optimum accuracies. Due to the limitations that unimodal systems suffer from, such as noisy data, non-universality, and susceptibility to spoof attacks, multibiometric systems have gained much interest in the research community on the grounds that they alleviate most of these limitations and are capable of producing better accuracies and performances. A multibiometric system combines two or more biometric sources in order to overcome their unimodal system counterparts and achieve higher accuracies. One of the important steps to reach this purpose is the choice of the fusion techniques utilized. A thorough study is performed to investigate the different fusion rules and schemes. In this work, a modeling step based on a hybrid algorithm that includes social rules derived from the swarm intelligence, Particle Swarm Optimization, and the concepts of natural selection and evolution, Genetic Algorithm, is used to combine the two modalities at the score level. This optimization algorithm is employed to select the optimum weights associated to the modalities being fused. The performance of the hybrid GA-PSO is compared to those of classical combination rules. For that purpose, the proposed schemes are experimentally evaluated on publicly available score databases (XM2VTS, NIST and BANCA) which come in clean and degraded conditions. An analysis of the results is carried out on the basis of comparing the techniques' resulting EER accuracies and ROC curves. Furthermore, the execution speed of the hybrid approach is compared to that of the single optimization algorithms GA and PSO