How Wi-Fi Signals Reveal Your Every Move: The Graph AI Breakthrough in Passive Sensing

Switzerland, Source:

Your body leaves an invisible trail as it moves through the air. This trail disturbs the radio waves that fill your home, creating a unique signature that standard Wi-Fi routers can detect. This technology, known as Channel State Information or CSI sensing, turns everyday internet equipment into passive motion detectors. It offers a privacy-friendly alternative to cameras and wearables by recognizing activities like walking or gesturing without requiring any special devices on the user.

The path to reliable detection has not been smooth. Different Wi-Fi hardware interprets signal data in varying ways, making it difficult for models to recognize patterns consistently across different devices. Multipath effects further complicate matters by creating complex relationships between signal frequencies that shift based on where you stand. Previous attempts often failed when conditions changed because they relied on rigid assumptions about how signals behave.

A new architectural approach addresses these inconsistencies by using multiresolution multiplex graphs. The process starts by resampling valid subcarriers to a common coordinate system, ensuring that input data remains consistent regardless of the hardware used. Two specialized encoders then work together to analyze the signal. One captures changes in amplitude over time, while the other models the propagation of signed subcarrier data. This graph component identifies both co-varying and counter-varying trajectories, capturing nuanced motion dynamics with greater precision than earlier methods.

Testing on the comprehensive CSI-Bench dataset highlights the model's strong generalization capabilities. It achieved a weighted-F1 score of 95.96% during in-domain testing and maintained high accuracy across different devices, environments, and users. This performance significantly surpasses baseline architectures such as LSTM, Transformer, and ResNet-18. The ability to function effectively across heterogeneous setups suggests readiness for practical applications in smart homes, healthcare monitoring, and security systems.

Constrained temporal analysis and gated relational conditioning enable the network to distinguish subtle activity patterns even when acquisition parameters change significantly. While the complexity of graph operations increases computational costs compared to simpler models, inference latency remains manageable on modern GPU hardware. This balance between high accuracy and practical deployability positions advanced graph-based representations as a key enabler for ubiquitous, passive activity recognition in connected environments.