Design of a knowledge distillation network for wifi-based indoor localization

The main purpose of indoor localization is to precisely locate users and help them navigate within an indoor area, like a building or campus, where GPS and other satellite technologies lack precision. Our methodology for achieving indoor localization has been to implement classifiers that use Received Signal Strength Indicator (RSSI) values of WiFi signals collected from smart hand-held devices. However, these RSSI values keep varying, often appreciably, from time to time and device to device. So, to instill more generalizability into the location prediction process, ensemble models have been built that can learn from the pros and cons of all of their member classifiers. In this paper, we have presented several neural network based ensemble models to compensate for the lack of detailed studies with ensemble models (especially neural network based ones) on indoor localization. Our second contribution lies in designing a knowledge distillation framework for the ensemble models that preserves the classification performance while make the system real-time responsive as the lightweight distilled model could be executed locally on the edge devices. Our proposed knowledge distillation framework distils the knowledge of a large neural network based ensemble classifier into a much smaller compressed classification model while maintaining the performance. We have implemented and shown the workings of the proposed knowledge distillation framework on three publicly available benchmark datasets. The proposed model have been found to achieve 83.95%, 93.10% and 96.48% accuracy for DataSet1, DataSet2 and DataSet3, respectively.