PRIVAtE: Passive Radar Interpretability using Variational Auto Encoders

Marco Cominelli, Paolo Braca, Leonardo M. Millefiori, Lance Kaplan, Mani Srivastava, Francesco Gringoli, Federico Cerutti2023 · DOI

Abstract

This paper presents PRIVAtE, a novel framework designed to enhance the interpretability of human activity recognition using Wi-Fi signals as passive radar. By leveraging Variational Auto-Encoders (VAEs), the system maps complex Channel State Information (CSI) into a low-dimensional latent space. We introduce a statistical decision function based on the Large Deviations Principle to provide human analysts with quantifiable error probabilities, bridging the gap between high-performance deep learning and actionable situational awareness.

Introduction

Modern security monitoring increasingly demands "device-free" sensing—detecting human presence and activity without invasive cameras or wearable trackers. Wi-Fi signals offer a ubiquitous solution; as individuals move through a space, they disturb the wireless field, leaving unique signatures in the Channel State Information (CSI). However, raw CSI data is high-dimensional and noisy, often resulting in "black-box" models that lack the transparency required for high-stakes environments. The PRIVAtE method addresses this by providing a statistically grounded, interpretable pipeline for activity recognition.

The PRIVAtE Framework

The core of the PRIVAtE method is a three-stage pipeline designed to transform raw signal interference into interpretable intelligence.

The PRIVAtE architecture: CSI from four antennas is processed by individual VAEs before classification by an MLP.
  1. Compression via VAEs: Raw CSI magnitude from four receiver antennas is fed into four independent VAEs. These encoders compress the noisy data into a simplified "electromagnetic signature" within a latent space.
  2. Classification: A Multi-Layer Perceptron (MLP) processes these latent variables to identify specific activities, such as walking, running, or jumping.
  3. Statistical Interpretability (D3F): To provide trust, the framework employs a Data Driven Decision Function (D3F). By applying the Large Deviations Principle, the system calculates the mathematical "distance" between activity distributions in the latent space, allowing for the estimation of Type I and Type II error probabilities.

Experimental Setup

The framework was validated using a public dataset collected with commercial Asus RT-AX86U routers in a 45m² indoor environment. One router transmitted 150 frames per second, while a second monitor router captured CSI across four antennas. The dataset includes five distinct classes: walk, run, jump, sit, and empty room.

Results and Statistical Interpretability

The PRIVAtE architecture achieved a test accuracy of approximately 95%. Beyond raw accuracy, the D3F analysis provided deep insights into the model's decision-making process.

Empirical density of the D3F output for 'Empty vs. Jump' and 'Empty vs. Run', showing clear statistical separation.

As shown in the density plots, active movements like jumping and running are highly separable from an empty room in the latent space. However, the analysis also highlighted challenges: "Run" and "Walk" were found to be the least separable active classes, and "Sitting" often overlapped significantly with the "Empty Room" class. This quantitative measure of uncertainty allows analysts to understand exactly when and why the system might fail, which is critical for real-world deployment.

Conclusion and Future Work

PRIVAtE demonstrates that VAEs can effectively distill complex Wi-Fi interference into interpretable features. By providing a mathematical framework for error analysis, it moves Wi-Fi sensing from a black-box prediction tool to a reliable source of situational awareness. Future research will focus on Neuro-symbolic AI to improve environmental generalization and the development of new loss functions based on statistical error probabilities to further refine model reliability.

Summary updated

Abstract

This paper aims to present a method to increase interpretability human analysts can have in actionable intelligence from analysing Wi-Fi signals used as passive radar systems for situational understanding. The Passive Radar Interpretability using Variational Auto Encoders (PRIVAtE) method is demonstrated using a recent dataset that estimates the latent distributions of antennas of the same Wi-Fi receiver to perform human activity recognition. The performance theoretical analysis of machine learning binary classification includes error probabilities using a statistical test based on a classification learned in the training phase. Results demonstrate that the compressed data obtained using the Variational Auto-Encoder is statistically very informative for providing situational understanding.
BibTeX
@misc{cominelli2023private,
  author = {Marco Cominelli and Paolo Braca and Leonardo M. Millefiori and Lance Kaplan and Mani Srivastava and Francesco Gringoli and Federico Cerutti},
  title = {{PRIVAtE: Passive Radar Interpretability using Variational Auto Encoders}},
  year = {2023},
  doi = {10.1109/techdefense59795.2023.10380847},
}
Figure