Insights From Inside: Toward Explainable WiFi Sensing

Abstract

WiFi sensing relies heavily on blackbox machine learning [Machine Learning (ML)] models due to the large feature space and complexity. Despite achieving very high accuracies in complex scenarios, the blackbox nature of these ML-based sensing techniques is commonly criticized. This is in fact a major source of mistrust as these models provide very little explanation supporting their decision, while often handling critical applications (e.g., elderly monitoring). In this paper, we investigate explainable artificial intelligence [eXplainable Artificial Intelligence (XAI)] techniques to shed light on the decisions and behaviors of such blackbox models. Specifically, we propose eXSense, a workflow designed based on state-of-the-art XAI techniques to analyze the behavior of blackbox models both locally and globally. To demonstrate its potential, we conduct an extensive analysis on two case studies from the recent sensing literature. Finally, leveraging the insights obtained from our analysis, we propose and evaluate changes to these models, thus enhancing their efficiency and reliability. This includes reducing the feature space by at least 80% with no/minimal loss (≤1%) to the model accuracy.