Artificial Bee Colony Hyperparameter Optimization of Convolutional Neural Networks for Multimodal Biometric Authentication in Smart-Home IoT Environments

Convolutional neural network (CNN) performance is highly dependent on model hyperparameters, yet manually selected configurations may not provide the best balance of recognition accuracy, error rate and computational cost for multimodal biometric authentication. This study evaluated the effect of Artificial Bee Colony (ABC) hyperparameter optimization on a CNN classifier for face-palm-iris recognition in a smart-home Internet of Things setting. The MULB dataset contained 10,266 biometric samples from 219 identity classes. Images were standardized to 128 × 128 pixels, transformed into a common 150-dimensional fused PCA-LDA representation, and divided into 8,212 training and 2,054 testing samples. A conventional CNN served as the baseline, while ABC searched convolutional filters, kernel size, dense units, dropout rate and learning rate. Each classifier was executed four times and assessed using accuracy, precision, recall, F1-score, false positive rate (FPR), false negative rate (FNR) and classifier-level recognition time. The ABC-CNN increased mean accuracy from 80.92 ± 0.30% to 84.81 ± 0.37% and F1-score from 80.66 ± 0.29% to 84.59 ± 0.38%. FPR fell from 0.0878 ± 0.0011% to 0.0694 ± 0.0012%, while FNR fell from 14.747 ± 0.12% to 10.85 ± 0.25%. Recognition time increased by 0.000062 s/sample. Paired-samples t-tests reported significant differences for all seven metrics (p < 0.001). The findings show that ABC-based optimization improved recognition effectiveness and reduced classification errors, with a small increase in classifier-level processing time.