Continuous Behavioral User Authentication on Mobile Devices using Online Machine Learning

Astract:

As smartphones provide more and more functionalities these days, the amount of sensitive information stored on these increases. Constant reauthentication requests for standard methods as pins annoy many users, who as a result tend to only insufficiently secure their phones. Additionally, authentication methods based on content only are easy to trick, leaving the phone unsecured once been passed. Continuous authentication based on behavioral profiles does not suffer from these drawbacks. The user gets continuously identified based on his interactions with the phone, which gets locked on suspicious behavior.This study is based on touch and sensor data gathered in a 30 minute experiment from 28 users. We first evaluate usefulness of temporal, spatial and sensor features to discriminate users according to their behavior. Based on these results, we use online machine learning to continuously identify users. While online learners with unlimited budget are known to suffer from unlimited growth in model size, our approach aims at bounding memory footprint by means of online learning on a budget. We evaluate how bounded model size affects classification accuracy compared to a unlimited environment.Our results show that it is possible to correctly classify anomalous behavior on the phone using online learning while dealing with limited model size. In case of keystroke dynamics, we were able to achieve comparable results to unlimited learners while working in a resource-constrained environment.