Dynamic Exit Selection for Comprehensive and Energy Efficient Gait-Based User Authentication on IoT Devices

In this study, we introduce Apollo, a dynamic and efficient gait-based user authentication system for Internet-of-Things (IoT) devices, which leverages the dynamic Early-Exit Neural Networks (EENets). As IoT devices proliferate across various applications, the need for effective authentication methods that cater to their resource constraints becomes crucial. Traditional authentication techniques, while robust, often fall short due to their high resource demands, making them unsuitable for IoT contexts. Gait-based authentication emerges as a viable solution, offering a non-intrusive and secure method. However, the implementation of gait-based authentication faces challenges, particularly due to the limited computational resources and stringent energy constraints of IoT devices. Our framework addresses these challenges by orchestrating Dynamic EENets, intermittent computing, and energy harvesting, enabling authentication with minimal or even negative energy consumption. Our approach is distinguished by the use of dynamic exit selection for EENets, which are lightweight, energy-aware, and capable of dynamically selecting the most appropriate time to exit based on current computational and energy resources. This adaptability ensures efficient authentication, reducing the overall computational load. Additionally, to overcome energy limitations, Apollo incorporates piezoelectric energy harvesting from user movements, further enhancing its viability for IoT applications. Our experimental results underscore Apollo's effectiveness, demonstrating over 85% accuracy in user authentication while significantly reducing computational demands by more than 34%. This blend of dynamic exit strategies, energy harvesting, and intermittent computing not only highlights the novelty of our research but also sets a new benchmark for gait-based authentication in IoT devices.