The literature is rich in techniques and methods to perform Continuous Authentication (CA) using biometric data, both physiological and behavioral. As a recent trend, less invasive methods such as the ones based on context-aware recognition allows the continuous identification of the user by retrieving device and app usage patterns. However, a still uncovered research topic is to extend the concepts of behavioral and context-aware biometric to take into account all the sensing data provided by the Internet of Things (IoT) and the smart city, in the shape of user habits. In this paper, we propose a meta-model-driven approach to mine user habits, by means of a combination of IoT data incoming from several sources such as smart mobility, smart metering, smart home, wearables and so on. Then, we use those habits to seamlessly authenticate users in real time all along the smart city when the same behavior occurs in different context and with different sensing technologies. Our model, which we called WoX+, allows the automatic extraction of user habits using a novel Artificial Intelligence (AI) technique focused on high-level concepts. The aim is to continuously authenticate the users using their habits as behavioral biometric, independently from the involved sensing hardware. To prove the effectiveness of WoX+ we organized a quantitative and qualitative evaluation in which 10 participants told us a spending habit they have involving the use of IoT. We chose the financial domain because it is ubiquitous, it is inherently multi-device, it is rich in time patterns, and most of all it requires a secure authentication. With the aim of extracting the requirement of such a system, we also asked the cohort how they expect WoX+ will use such habits to securely automatize payments and identify them in the smart city. We discovered that WoX+ satisfies most of the expected requirements, particularly in terms of unobtrusiveness of the solution, in contrast with the limitations observed in the existing studies. Finally, we used the responses given by the cohorts to generate synthetic data and train our novel AI block. Results show that the error in reconstructing the habits is acceptable: Mean Squared Error Percentage (MSEP) 0.04%.
