Task-Based Continuous Authentication Using Wrist-Worn Devices

Activity-based continuous authentication methods have been shown to be effective for identifying individual users. Existing classification-based approaches typically learn models of activities of daily living (ADLs) using sensor data from an individual's mobile device. For activity-based authentication to be widely applicable, we contend that such approaches should also consider activities that are aligned with an task-related activities (e.g., picking packages in a shipping warehouse), which are typically shorter and burstier than ADLs. In this paper, we explore the feasibility of task-driven continuous authentication and implement a pipeline of machine learning approaches that capture task-based continuous authentication models. In a scenario-based evaluation using real-world ubiquitous wrist-worn sensor data, our approach can detect distinct users using task-specific models with 94% accuracy on average.