Touch-Based Continuous Mobile Device Authentication Using One-vs-One Classification Approach

In recent years, there has been a significant increase in the number of users handling personal information on their smartphones. Although smartphones offer standard login authentication to safeguard the information, there are inherent risks of malicious users hacking these authentication mechanisms. Given this context, touch-based authentication using touch gestures has gained attention as a post-login authentication method. Typically, touch-based authentication research identifies the device owner as the legitimate user and all others as illegitimate users, employing the One-vs-Rest (OvR) classification approach. However, due to the variability in user touch patterns, training classifiers using data from multiple users for the rest class can be challenging. To address this issue, we propose a new classification method that adopts the One-vs-One (OvO) classification approach to improve accuracy. In the OvO approach, each classifier is trained using data from one legitimate user and one illegitimate user, with subsequent user identification performed through ensembling a set of OvO classifiers' output. We conducted an experimental evaluation to compare the performance of the OvR and the OvO approaches across six scenarios and five classification models. Our results showed that the proposed OvO approach outperformed the OvR approach in terms of average equal error rate (EER) among 120 combinations of scenarios and models. The best-performing OvO approach, combined with dynamic classifier selection, achieved the lowest EER of 1.42%, proving its high classification performance.