Application of Mixture Density Network for Sample Generation in Behavioral Biometrics

This article presents the results of research on the development of a biometric system for identifying individuals based on their gait using smartphones. Experiments were conducted on a publicly available gait database of 13 individuals, containing data collected over three different days (enabling so-called cross-day validation). A CNN classifier with an attention mechanism was used as the decision-making module in the study, with segmented gait cycles input. The article investigated the impact of enriching the training set with artificially created synthetic samples. The influence of data generated by LSTM-MDN architecture models was compared with competing methods: Riemannian Hamiltonian VAE (available in the PyRaug python module) and timeVAE. The use of synthetic samples generated by the LSTM-MDN model enabled an increase in the biometric system’s effectiveness from F1-score of 0.898 to 0.939.