Evaluating multi-modal mobile behavioral biometrics using public datasets

Abstract

Behavioral biometric-based continuous user authentication is promising for securing mobile phones while complementing traditional security mechanisms. However, the existing state of art perform continuous authentication to evaluate deep learning models, but lacks examining different feature sets over the data. Therefore, we evaluate the performance of user authentication based on acceleration, gyroscope (angular velocity), and swipe data from two public mobile datasets, HMOG (Hand-Movement, Orientation, and Grasp) (Sitová et al., (2015) dataset et al. (2015)) and BB-MAS (Behavioral Biometrics Multi-device and multi-Activity data from Same users) (Belman et al., (2019) dataset et al. (2019)) extracted with different feature sets to observe the variation in authentication performance. We evaluate the performances of both individual modalities and their fusion. Since the swipe data is intermittent but the motion event data continuous, we evaluate fusion of swipes with motion events that occur within the swipes versus fusion of motion events outside of swipes. Moreover, we extract Frank et al.’s (2012) Touchalytics features Frank et al. (2012) on the swipe data but three different feature sets (median, HMOG (Sitová et al. (2015)), and Shen’s (Shen et al. (2017))) on the motion event data, among which the Shen’s features were shown to perform the best. More specifically, we perform score-level fusion for a single modality utilizing binary SVMs (Support Vector Machine). Furthermore, we evaluate the fusion of multiple modalities using Nandakumar’s likelihood ratio-based score fusion (Nandakumar et al. (2007)) by utilizing both one-class and binary SVMs. The best EERs (Equal Error Rates) of fusing all three modalities when using the one-class SVMs are 8.8% and 0.9% for HMOG and BB-MAS respectively. On the other hand, the best EERs in the case of binary SVMs are 1.5% and 0.2% respectively. Observing the better performances of BB-MAS compared to HMOG in swipe-based experiments, we examine the difference of swipe trajectory between the two datasets and find that BB-MAS has longer swipes than HMOG which would explain the performance difference in the experiments.