MoveAR: Continuous Biometric Authentication for Augmented Reality Headsets

Augmented Reality (AR) headsets are rapidly coming to consumer and professional markets. The lack of traditional input interfaces for these devices motivates the need to research novel methods of achieving security primitives such as user authentication. Given the various inertial sensors that the headsets use to position users in their environment, we propose, investigate, and evaluate the potential for a continuous biometric authentication system based on the distinct ways in which people move their heads and interact with their virtual environments. We collect samples of the spatial and behavioural patterns from a group of users wearing an AR headset. Using this data, we propose a multitude of novel models and machine learning pipelines that learn the unique signature of AR users as they interact with the virtual environment and AR objects. We evaluate multiple supervised machine learning algorithms, including k-Nearest Neighbours, Random Forest, Support Vector Machine (SVM), and Bag of Symbolic-Fourier-Approximation Symbols (BOSS) for two different sets of input data and parameters. We achieve a balanced accuracy score of 92.675% and an EER of 11% using an Adaptive Boost Random Forest classifier together with our proposed series of novel, AR-specific preprocessing methods used on our current dataset. This demonstrates that it is indeed possible to profile and authenticate AR head-mounted display users based on their head movements and gestures.