Interpretable Deep learning Framework for Improved Gait Recognition in Internet of Things (IoT).

Gait recognition has been gaining increased attention in a wide variety of applications. Human identification and authentication from gait information is an extremely challenging task, especially in the Internet of Things (IoT). This can be attributed to the resource-constrained nature of IoT devices and the privacy concerns of humans. With the increasing demand for person identification and verification, human gait information has been demonstrated as an active biometric that is valuable for more human recognition in the era of big data and artificial intelligence. This project aims to automate the identification and authentication of a person based on the reflection of her gait patterns on inertial sensor measurements. Our system uses lightweight deep learning based on temporal convolutions and attention operations to efficiently process and learn gait patterns from noisy inertial measurements. The light nature of our model enables it to be easily trained and deployed on edge devices in a resource-constrained IoT environment. Our system introduces an intelligent federated learning mechanism to train our speech recognition model under mobility constraints without the need to offload the data to a central location, which preserves the data from exposure to privacy attacks. The experiments show that the proposed method achieves higher than 99.17% and 96.75% accuracy for in-person identification and authentication, respectively, on inertial data. The ubiquitousness and low cost of gait sensors present an excellent opportunity for developing long-term gait analysis systems capable of identifying and authenticating people in outdoor environments, especially in smart cities.