Efficient feature selection for electroencephalogram-based authentication

Opposed to classic authentication protocols based on credentials, biometric-based authentication has recently emerged as a promising paradigm for achieving fast and secure authentication of users. Among the several families of biometric features, electroencephalogram (EEG)-based biometrics is considered as a promising approach due to its unique characteristics. Classification systems based on machine learning allow processing of large amounts of data and performing accurate attribution of each signal to the most relevant group, thus representing an invaluable tool for EEG-based biometrics. This paper provides an experimental evaluation of the performance achievable by EEG-based biometrics employing machine learning. We consider several groups of EEG signals and propose a suitable feature extraction criterion. Then, the extracted features are used along with neural network-based classification algorithms, K Nearest Neighbours (KNN), and eXtreme Gradient Boost (XGBoost) for attributing any EEG signal to a subject. A full feature set and a reduced feature sets are considered and tested on three public data sets. The feature selection criteria are based on a correlation map among features, ANOVA F-test, and logistic regression weights. The results show that the reduced feature sets achieves a significant reduction in computation time over the full feature set, while also providing some improvement in performance.