🤖 machine learning

ARGUS: Attention-Guided Transformers for Scalable Person Identification Using Wi-Fi Telemetry

The paper presents Argus, a scalable, passive Wi-Fi person identification system that achieves high accuracy on large-scale datasets by converting Channel State Information into compact statistical maps ("statgrams") processed by an efficient, attention-guided Transformer architecture, significantly reducing computational costs compared to raw-CSI baselines.

Original authors: Nayan Sanjay Bhatia, Pranay Kocheta, Yuhan Li, Katia Obraczka

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nayan Sanjay Bhatia, Pranay Kocheta, Yuhan Li, Katia Obraczka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine walking into a room where the air itself is constantly humming with invisible radio waves, bouncing off walls, furniture, and the people inside. These waves are the same ones that carry your Wi-Fi signal, connecting your phone to the internet. For years, scientists have known that when a person moves through this invisible field, their body subtly distorts the waves, creating a unique fingerprint in the signal. This idea has sparked a race to build identification systems that work without cameras or wearable devices, offering a way to recognize people simply by how they interact with the air around them. However, turning these fleeting, noisy distortions into a reliable identity has been a stubborn challenge. The signals are messy, they change depending on the shape of the room, and they often require a person to be moving in specific ways to be detected.

A team of researchers has now introduced a new system called Argus that tackles these problems by changing how the computer "looks" at the data. Instead of trying to process every single burst of radio information as it happens, Argus first condenses the raw signal into a compact, statistical map. Think of this map as a summary sheet that captures the essential shape of the signal's behavior over a few seconds, stripping away the noise and redundancy. The system then uses a specialized type of artificial intelligence, designed to read these maps like a series of small image patches, to identify who is in the room. The researchers tested this approach on a dataset of 154 different people, asking the system to identify them while they stood still or performed simple, stationary activities. The results were striking: by combining the evidence from multiple short windows of time, the system correctly identified the right person in nearly 85 percent of cases, and it was able to narrow the possibilities down to the top five candidates in more than 99 percent of cases.

What makes this approach particularly significant is not just its accuracy, but its efficiency. Previous methods that tried to analyze the raw, unprocessed radio waves required massive amounts of computing power and often failed as the observation time grew longer. In contrast, Argus achieves its high accuracy while using roughly twenty-seven times less computing power than the strongest competing systems. This efficiency comes from the initial step of creating the statistical map, which allows the computer to ignore vast amounts of irrelevant data. The researchers found that the system does not need to see the entire history of the signal to make a decision; it only needs to focus on a few key parts of the summary map that contain the most useful information. In fact, they discovered that they could remove more than half of the data points from the map without significantly hurting the system's ability to identify people, proving that the most important details are concentrated in specific areas.

The study also tested how well this system works when the environment changes, such as moving from one room to another or switching between different Wi-Fi frequencies. Here, the system revealed its current limits. While it performed very well when identifying people in the same room where it was trained, it struggled significantly when asked to recognize those same people in a completely different room without any prior training in that new space. This suggests that the unique way radio waves bounce off the walls of a specific room is a dominant factor, often overpowering the subtle signals created by the person themselves. The researchers also noted that the system is not yet perfect at telling the difference between a known person and a stranger; it is better at creating a short list of likely candidates than at making a final, definitive identification on its own.

Despite these limitations, the work demonstrates a clear path forward for passive identification. By showing that a compact summary of the signal is more powerful than raw data, the researchers have provided a blueprint for building systems that are both accurate and practical for real-world use. The system does not require people to wear special devices or perform specific movements, making it a promising candidate for applications where privacy and convenience are paramount. However, the researchers are careful to note that before such technology can be widely deployed, it must be paired with stronger safeguards to ensure it cannot be tricked by unknown individuals and that it respects user consent. The future of this technology lies not in making the system more complex, but in refining how it interprets the quiet, invisible language of the air around us.

Technical Summary: ARGUS – Attention-Guided Transformers for Scalable Person Identification Using Wi-Fi Telemetry

Problem Statement

Accurate and efficient person identification is critical for access control and personalized services, yet existing biometric methods face significant limitations. Camera-based systems introduce privacy concerns and are sensitive to spoofing, occlusion, and environmental conditions, while wearable-based approaches require user compliance. Wireless alternatives using commodity Wi-Fi Channel State Information (CSI) offer a promising, low-cost, and passive solution. However, current CSI-based identification systems struggle with several challenges:

Data Complexity: Raw CSI streams are noisy, high-dimensional, and suffer from synchronization-induced phase offsets that require calibration.
Domain Shift: CSI signals are highly dependent on room geometry, hardware, and device placement, making cross-domain generalization difficult.
Motion Dependency: Many existing systems rely on gait or deliberate movement, limiting their utility for identifying stationary subjects.
Scalability and Efficiency: Transformer-based methods targeting resting subjects have been evaluated primarily on small cohorts, and closed-set accuracy alone fails to address open-set rejection or computational efficiency in large-scale deployments.

Methodology

The authors propose Argus, a passive Wi-Fi sensing system designed to identify individuals from commodity CSI without attached devices or prescribed motion. The system architecture consists of three primary components:

Statgram Representation:
Instead of modeling raw packet streams directly, Argus converts short CSI windows into compact statistical maps called statgrams.
- Preprocessing: Raw complex CSI ( $H$ ) undergoes robust normalization (using median and interquartile range) and phase unwrapping with linear detrending to remove synchronization offsets.
- Feature Extraction: The system generates multiple "views" of the signal: Amplitude ( $A$ ), Calibrated Phase ( $P$ ), and Phase-Delta ( $D$ ).
- Statistical Aggregation: For each view, statistics (mean, standard deviation, percentiles, energy, and frequency-domain features via real-FFT) are computed over time and grouped subcarriers. These are stacked into a tensor $S \in \mathbb{R}^{C \times R \times W}$ , where $C$ is the number of views, $R$ is the number of statistic rows, and $W$ is the width determined by capture geometry.
Patch Transformer Architecture:
- The statgram is divided into coarse rectangular patches, which are flattened and linearly projected into tokens.
- A learnable CLS token is appended to the sequence.
- A lightweight decoder-only Transformer processes these tokens. The CLS token, positioned last, attends to all preceding patches via causal attention.
- Inference: For a physical recording, logits from overlapping windows are aggregated (averaged) to produce a final segment-level prediction. For single-person identification, a softmax head is used; for multi-user scenarios, independent binary logits are employed.
Attention-Guided Compression:
Argus utilizes the trained model to rank the importance of statgram patches. By occluding patches with training-set mean values and measuring the drop in prediction confidence, the system identifies which regions contain identity-relevant information. This allows for input compression (retaining only top-ranked patches) without significant accuracy loss.

Key Contributions

System Design and Evaluation: The paper details the Argus pipeline and evaluates it on two datasets: the EHealth dataset (154 participants, stationary and limited motion) and the WiMANS dataset (multi-user, multi-room, 2.4/5 GHz bands).
Compact Representation vs. Raw Context: The authors demonstrate that compact statgram representations are more effective than simply extending raw temporal context. On the EHealth dataset, Argus improves Top-1 accuracy by 7.75 points over a raw-CSI Transformer baseline at 60-second evidence while using 4.4× fewer FLOPs per window.
Attention-Occlusion Ranking: A validation-only method is introduced to rank statgram patches. Retaining only the top 4 of 8 patches on EHealth preserves full single-window accuracy, and retaining 10 of 15 patches on WiMANS incurs only a 0.7-point accuracy drop.
Realistic Deployment Limits: The study explicitly evaluates open-set rejection and cross-environment transfer, revealing that while Argus performs well in closed-set scenarios, zero-shot transfer across rooms is weak, and open-set rejection remains moderate.

Results

EHealth Dataset (154 Subjects):

Accuracy: Argus achieves 78.88% ± 1.62% Top-1 accuracy on 6-second windows. When aggregating 19 overlapping windows (60 seconds), Top-1 accuracy rises to 84.85% ± 1.31%, with Top-5 reaching 99.26%.
Efficiency: The model requires 33.6 MFLOPs per window, compared to 149 MFLOPs for the competing THAT baseline.
Ablation: In the single-radio setting, amplitude channels carry the majority of discriminative signal; phase-only models perform poorly without calibration.

WiMANS Dataset (Multi-User, Multi-Room):

Accuracy: Argus achieves a mean exact-match accuracy of 95.07%, remaining within 1.23 percentage points of the strongest per-configuration baselines on average.
Efficiency: Argus uses 0.061 GFLOPs per recording, which is 27× fewer inference FLOPs than the THAT baseline. With patch pruning, this speedup increases to 48×.
Transfer Learning: Zero-shot transfer across rooms fails (14–28% accuracy). However, few-shot mixed training (using source data + small target samples) significantly improves performance, reaching 82–93% accuracy with 25% of the target data.

Significance and Claims

The paper claims that compact CSI statistics can enable scalable passive identification, offering a viable alternative to camera and wearable biometrics. The significance of Argus lies in its ability to:

Scale: Successfully identify individuals in large cohorts (154 subjects) and multi-user environments without requiring gait or movement.
Efficiency: Drastically reduce computational requirements through statgram compression and lightweight Transformers, making deployment on edge devices feasible.
Interpretability: Use attention mechanisms to identify and prune non-essential signal regions, providing insights into which parts of the CSI signal drive decisions.

However, the authors maintain a modest stance regarding deployment readiness. They emphasize that open-set rejection is currently insufficient for standalone biometric gating (unknown users may be confidently misidentified), and cross-environment generalization remains a significant barrier due to room-specific multipath effects. Consequently, Argus is positioned as a high-recall shortlist generator or a component within a larger verification system, rather than a standalone solution for all identification scenarios. The paper also highlights the necessity of governance, including informed consent and local processing, for the ethical deployment of passive biometric sensing.