Meta-learning for selection of adaptation strategies in a biometric system
Users are constantly entering personal information on cellphones and computers to access applications that are not always secure, which can lead them to be targets of malicious attacks and serious losses. Some authentication systems have started to add biometric recognition as an extra layer of security. For example, user and password-based access systems can have an extra layer of security based on keystroke dynamics. Keystroke dynamics is a behavioral biometric modality capable of recognizing a user based on their typing rhythm on a keyboard. This modality has a low implementation cost and is not very intrusive for users, which can increase acceptance. However, the biometric data of the same user may change over time, which may lead to lower performance of the biometric system. A promising way to address this problem is through adaptive biometric systems, which can automatically update users’ biometric references over time. Some studies have shown that defining specific adaptation strategies for each user can improve system performance. But despite this, there is little work in the literature studying how meta-learning approaches can help solve the problem of choosing the adaptation strategy for users of a biometric system based on keystroke dynamics. To fill this gap in the literature, this work proposes a new way of structuring adaptive biometric systems, called meta-adaptive biometric system. In this system, meta-learning techniques are used to assist in the task of automatically recommend the best adaptation strategy for a user.