Biometric Recognition Using Multimodal Physiological Signals

In this paper, we address the problem of biometric recognition using the multimodal physiological signals. To this end, four different signals are considered: heart rate (HR), breathing rate (BR), palm electrodermal activity (P-EDA), and perinasal perspitation (PER-EDA). The proposed method consists of a convolutional neural network that exploits mono-dimensional convolutions (1D-CNN) and takes as input a window of the raw signals stacked along the channel dimension. The architecture and training hyperparameters of the proposed network are automatically optimized with the sequential model-based optimization. The experiments run on a publicly available dataset of multimodal signals acquired from 37 subjects in a controlled experiment on a driving simulator show that our method is able to reach a top-1 accuracy equal to 88.74% and a top-5 accuracy of 99.51% when a single model is used. The performance further increases to 90.54% and 99.69% for top-1 and top-5 accuracies, respectively, if an ensemble of models is used.