Classification of Electrocardiography Signals for User Authentication Based on Ensembles with Random Undersampling

Inadequate authentication processes constitute a major security flaw in several applications. In mobile healthcare, they can lead to patients not receiving proper attention, or to delays that might result in death. Towards a safe and fast authentication with little or no user intervention, modern systems have explored physiological traits for biometric recognition, among which electrocardiography (ECG) signals offer some important advantages, including subject specificity and pattern robustness with no conscious effort or behavioral change. This paper explores ensembles for the classification of ECG signals for authentication. Given its robustness to dataset imbalances, Random Under-Sampling Boosting (RUSBoost) strategy, not yet explored for such an application, to the best of our knowledge, was applied. Its performance was compared with that of Nearest Neighbour Search, which is the basis of some previous efforts. A cross-validation procedure, used for the comparison, employed random subsampling to define the training and testing datasets, and three objective metrics, namely accuracy, sensitivity, and F1-score were considered. The results indicated a better performance of RUSBoost regarding accuracy (97.4%), sensitivity (96.1%) and F1-score (97.4%).