Real-world continuous smartwatch-based user authentication

Abstract User authentication is often regarded as the “gatekeeper” of cyber security. It has, however, long suffered from significant usability issues that have resulted in research focussing upon frictionless and transparent biometric approaches. Activity-based user authentication—a technique that authenticates a user by what they are physically doing at a specific point in time has attracted significant attention, particularly due to the increasing popularity of smartwatches. This research aims to overcome limitations in prior work by exploring the viability of the approach in real-world conditions. The study presents two principal experiments, one focused upon a constrained environment to provide a control and a second reflecting real-life. With over 1000 h of sampled data across 60 participants, the study sought to explore sensor, feature composition, and classifier design to explore the practical viability of the approach. Whilst the control experiment achieved best case Equal Error Rate of 0.29%, an improvement upon the prior art using optimisation, the best-case real-world results were not too far behind at 0.7%. This demonstrates that whilst the feature generated in the real-life experiment are subject to increased levels of noise, the performance is viable within the context of a transparent and continuous user authentication approach.