Kailong Wang

With the wide deployment of smart environments and IoT devices, WiFi sensing has demonstrated its great convenience and contactless sensing capabilities in supporting a broad array of applications. However, designing a ubiquitous WiFi sensing system for heterogeneous scenarios in practice is still a big dilemma as the system performs poorly when the testing data is significantly different from the training data caused by domain variations. To address this dilemma, existing studies involve extra efforts to develop new features or even to retrain the original model under environmental variations. However, none of them can resolve the dilemma completely. In this work, we conduct a comprehensive study on the domain variation problem to make WiFi sensing robust and accurate in reality. Our definition of domains is comprehensive and includes environments, surrounding settings, user differences, user’s facing directions, user’s positions relative to WiFi sensors, and user participating time frames. Our innovation is to achieve reliable WiFi sensing across all the domains based on the conformal prediction framework. Our approach quantifies the conformity (i.e., similarity) between the testing WiFi samples and the training samples, then labels the testing samples with the most probable class(es). We develop a novel cross-domain transformal prediction scheme based on the multivariate kernel density estimation to effectively assess and learn the conformity of each domain in the training data. To meet various application-specific requirements, we further develop two approaches to fuse the knowledge of conformity derived from the training domains to perform predictions. Extensive experiments with both self-collected and public datasets show that our framework can improve prediction accuracies from 30% to 74% improvements in three most representative WiFi-based applications across six types of domain variations.