Development of a Multimodal Biometric Security System using Modified Convolutional Neural Network

Abstract

Biometrics is the biological measurement of physiological or behavioral attributes of humans. These characteristics are unique to each individual and remain unaltered throughout human lifetime. Several multimodal biometric security systems have been developed using convolutional neural network but few of them have been able to handle the challenges of complexities associated with CNN in terms of recognition rates and time. In this work, a multimodal biometric security system that uses a modified CNN (CNN-GA) for feature extraction and classification was developed. The System was tested on a database consisting of 1026 trained images and 684 probe images of face, ear and iris biometrics. The design, implementation and running/testing of the entire system were done on MATLAB R2016a programming platform. The multimodal images were first preprocessed. Feature extraction and classification were carried out using Convolution Neural Network-Genetic Algorithm (CNN-GA). It was the optimal classification result that was used to make final decision on whether to accept or reject probe images. The performance evaluation of the developed system was carried out using false positive rate, sensitivity, specificity, precision, recognition accuracy and recognition time. The result shows that at varying threshold values of 0.20, 0.35, 0.50 and 0.76, the CNN-GA outperforms the standard CNN as applied to the developed system in terms of sensitivity, specificity, precision, recognition accuracy and time. At the threshold value of 0.76, CNN-GA achieved a sensitivity of 97.66%, specificity of 98.25%, Precision of 99.40%, recognition accuracy of 97.81% and recognition time of 455.54Secs while the standard CNN achieved a sensitivity of 95.91%, specificity of 92.98%, Precision of 97.62%, recognition accuracy of 95.18% and recognition time of 565.02Secs.