WISDOM: A framework for scaling on-device Wi-Fi sensing solutions

Abstract

Recent innovations in Wi-Fi sensing capitalizes on a host of powerful deep neural network architectures that make inferences based on minute spatio-temporal dynamics in the wireless channel. Many of such inference techniques being resource intensive, conventional wisdom recommends offloading them to the network Edge for further processing. In this paper, we argue that edge based sensing is often not a viable option for many applications (due to cost, bandwidth, latency etc.). Rather, we explore the paradigm of on-device Wi-Fi sensing where inference is carried out locally on resource constrained IoT platforms. We present extensive benchmark results characterizing the resource consumption (memory, energy) and the performance (accuracy, inference rate) of some typical sensing tasks. We propose WISDOM, a framework that, depending on capabilities of the hardware platform and application’s requirements, can compress the inference model. Such context aware compression aims to improve the overall utility of the system — maximal inference performance at minimal resource costs. We demonstrate that models obtained using the WISDOM framework achieve higher utility compared to baseline models that are just quantized for 83% of the cases. While for non-compressed models it has higher utility 99% of the time.