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

Nayan Sanjay Bhatia; Pranay Kocheta; Yuhan Li; Katia Obraczka

Source-linked AI summary

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

Nayan Sanjay Bhatia, Pranay Kocheta, Yuhan Li, Katia Obraczka

arXiv:2608.14670v1Aug 2, 2026cs.LGcs.AIcs.CV

TL;DR

Device-free Wi-Fi identification remains limited by reliance on deliberate movement and small-scale evaluation. Argus uses compact CSI statgrams with a lightweight Transformer, achieving strong identification accuracy while reducing inference cost. Its deployment scope remains bounded by moderate open-set rejection and cross-environment barriers.

Problem
Existing CSI identification often relies on gait or deliberate movement, while Transformer approaches for resting subjects have mainly been evaluated on small cohorts.
Method
Argus converts short raw CSI windows into compact statgrams and processes their coarse patches with a lightweight decoder-only Transformer for identity prediction.
Results
84.85% ± 1.31% Top-1 and 99.26% Top-5 accuracy follow 60-second aggregation on 154 subjects, while WiMANS performance remains within 1.23 points of stronger baselines using 27× fewer inference FLOPs.
Takeaways & Limitations
Compact CSI statistics can support scalable passive identification, but practical deployment still requires stronger open-set rejection and cross-environment generalization.
Takeaways & Limitations
Moderate unknown-user rejection makes Argus better suited to shortlisting likely identities than serving as a standalone biometric gate.

Abstract

from arXiv · show

Passive, device-free person identification offers an alternative to camera- and wearable-based biometrics, yet existing wireless approaches rely largely on gait or activity cues and are rarely evaluated at scale. In this paper, we present \emph{Argus}, a passive Wi-Fi sensing system that identifies people from commodity Channel State Information (CSI) without requiring an attached device or a prescribed motion. Argus converts short CSI spans into compact \emph{statgrams}: statistical maps built from the channel views available on a given device. A lightweight decoder-only Transformer then reads coarse statgram patches as tokens, and segment-level logit aggregation combines evidence over time. On a 154-subject CSI dataset evaluated with a strict physical-segment split, Argus reaches $78.88\% \pm 1.62\%$ Top-1 accuracy on 6-second windows and $84.85\% \pm 1.31\%$ after aggregating 19 overlapping windows over a 60-second segment; Top-3 and Top-5 reach $98.61\%$ and $99.26\%$. For a 60-second statgram, Argus improves over a raw-CSI Transformer baseline by 7.75 points while using $4.4\times$ fewer FLOPs per window. Attention-guided compression preserves full single-window accuracy with only half of the EHealth patches. On WiMANS, a multi-user benchmark across three rooms and two Wi-Fi bands, Argus remains within 1.23 percentage points of the strongest per-configuration baselines on average while using $27\times$ fewer inference FLOPs. These results show that compact CSI statistics can scale passive identification while also exposing deployment limits in open-set rejection and cross-room transfer.

Introduction

Argus addresses the privacy, robustness, and deployment limitations of conventional person identification by using commodity Wi-Fi CSI for passive, device-free recognition. It compresses CSI into statistical maps processed by a decoder-only Transformer and evaluates this design across large-cohort and multi-user datasets.

Motivation: Commodity Wi-Fi CSI offers a low-cost alternative for person identification because it is ubiquitous in homes, offices, and public spaces.CSI captures attenuation, reflections, and phase variations influenced by the human body.
Challenges: CSI-based identification remains difficult because raw streams are noisy, high-dimensional, phase-offset, and sensitive to room geometry, hardware, and device placement.Existing systems also commonly depend on gait or other deliberate movement.
Argus design: Argus converts short raw CSI windows into compact statgrams that preserve separate signal views and summarize temporal and subcarrier variation with statistical features.The statistics include mean, standard deviation, percentiles, and energy.
Argus design: A decoder-only Transformer processes coarse statgram patches to perform passive person identification from Wi-Fi CSI telemetry.The system is introduced as a device-free approach that does not require prescribed motion.
Evaluation: 78.88% ± 1.62% Top-1 accuracy on 6-second windows increased to 84.85% ± 1.31% on 60-second segments, with Top-5 reaching 99.26% on the 154-subject EHealth dataset.Argus was also evaluated on WiMANS, a public multi-user benchmark spanning multiple environments.

Related Work

Prior wireless identification methods face practical constraints involving line of sight, privacy, hardware, motion dependence, environmental shift, and limited cohort scale. Recent work addresses adaptability and Transformer-based CSI modeling, but scaling across rooms and people remains difficult.

CSI-based person identification: Traditional vision and radar biometrics can achieve high accuracy but face line-of-sight, privacy, lighting, or specialized-hardware constraints.Radar avoids lighting dependence by analyzing RF reflections, with many frameworks exploiting micro-Doppler signatures.
CSI-based person identification: Cross-domain sensing and few-shot identity-similarity learning target Wi-Fi domain adaptability, but room-dependent channels and reliance on large body motion remain limitations.Physiological micro-motions can perturb wireless channels even when a person is at rest.
Transformers and interpretability: Transformers model long-range temporal dependencies and combine amplitude and phase perturbations, while recent dual-branch systems remain mostly limited to small cohorts.The passage identifies scale and environment shift as continuing challenges for CSI person identification.

Argus

Argus is a generalized CSI-based passive identification pipeline that converts short, noisy CSI spans into configurable statgrams and processes their patches with a lightweight decoder-only Transformer. It supports task-specific outputs, deterministic aggregation across overlapping windows, and validation-guided statgram patch compression.

Statgram construction: Argus can use amplitude, calibrated phase, or other stable CSI views, while its statgram dimensions remain configurable across radios, tasks, and capture geometries.The method does not assume fixed numbers of radios, antennas, channel bins, or enrolled users.
Statgram construction: Argus converts short CSI spans into statgrams to reduce outliers, packet jitter, hardware-dependent scaling, and redundant high-frequency variation.The representation is built from one or more channel views derived from time-ordered CSI measurements.
Transformer inference: Statgram patches become projected tokens for a lightweight decoder-only Transformer whose final CLS embedding feeds either a softmax or independent binary-logit output head.The CLS token is placed last so causal attention can still attend to every statgram patch.
Transformer inference: Segment-level inference averages logits across overlapping windows from the same physical recording, improving robustness without changing the trained model.This aggregation is deterministic.
Patch compression: Patch compression ranks statgram patches by validation-time true-class probability drops after mean-value masking, with accuracy drop as a secondary signal.Test-time compression is evaluated only after the validation ranking is fixed, alongside ranked mean-fill, ranked zero-fill, and random masking.

Experimental Methodology

Argus is evaluated on EHealth and WiMANS using dataset-specific protocols, preprocessing, splits, and multi-model comparisons. The methodology also includes open-set identification, unknown-user rejection, and calibration analysis on EHealth.

Datasets: 154 participants comprise EHealth, collected in a controlled 3 m × 4 m room across 17 standardized 60-second positions or activities.Fourteen activities were static, supporting passive identification evaluation across varied physical characteristics.
Datasets: 11,286 three-second WiMANS recordings span three room environments and two Wi-Fi bands, with recordings containing zero to five of six enrolled users.Complex CSI is re-extracted from raw traces to construct per-link amplitude and calibrated phase views; the resulting statgram uses 15 patch tokens plus CLS.
Evaluation Protocol: EHealth uses a subject-stratified physical-segment split with 11 training, 3 validation, and 3 test segments per participant, yielding 462 physical test segments.Neighboring windows from the same recording cannot cross splits, and demographic or physiological metadata is excluded.
Baselines: On EHealth, Argus is compared with THAT, a CNN-BiLSTM, and a full Argus encoder-decoder using the same statgram input and five seeds.The encoder-decoder has 4.2M parameters versus 1.9M for the decoder-only model.
Open-Set Evaluation: Open-set evaluation withholds 31 of 154 participant IDs, retrains on 123 enrolled IDs, and measures known-user accuracy, unknown rejection, and 15-bin ECE.Unknown rejection uses maximum softmax confidence and normalized negative entropy, with confidence-threshold sweeps.

Results

Argus achieves strong closed-set identification accuracy and efficiency on EHealth, with gains from temporal aggregation, compact statgrams, and attention-guided compression. Results on WiMANS show competitive multi-user recognition, while open-set rejection and cross-room transfer remain deployment limitations.

EHealth identification: 84.85% ± 1.31% Top-1 accuracy follows aggregation of 19 overlapping windows over 60 seconds, versus 78.88% ± 1.62% on a single 6-second window.Top-3 and Top-5 reach 98.61% and 99.26%.
EHealth identification: 7.75 points is Argus’s 60-second advantage over THAT, while per-window computation falls to 33.6 MFLOPs from 149 MFLOPs.A statgram pass takes 1.21 ms median on Apple Silicon, with approximately 23 ms for an x19 segment decision.
Ablations and compression: 79.72% ± 1.49% single-window Top-1 accuracy from amplitude-only input is statistically indistinguishable from the full model; phase-only and phase-delta-only reach 41.47% and 3.00%.The result makes phase calibration and channel selection important design choices.
Ablations and compression: 78.97% versus 78.88% preserves full single-window accuracy when retaining the top 4 of 8 ranked patches, at 18.6 MFLOPs per window.Accuracy declines only below 3 retrained patches, and random retention performs worse at the same budgets.
WiMANS results: 95.07% mean exact-match accuracy on WiMANS is within 1.23 points of the strongest per-configuration baselines across nine environment-band configurations.The remaining gaps concentrate in empty-room 5 GHz and empty-room dual-band settings.

Discussion and Future Work

The discussion identifies compact statgrams as Argus’s main contribution, while noting that longer raw temporal context does not necessarily improve performance. It also outlines limits in channel-view selection, open-set rejection, cross-environment transfer, and responsible deployment.

Contribution: Statgrams expose identification-relevant CSI structure through compact representations that remain learnable with short token sequences.This is presented as Argus’s main contribution.
Representation and aggregation: Longer raw temporal contexts can degrade raw-CSI Transformer performance, whereas statgram aggregation improves and ranked patch compression preserves most accuracy.The discussion contrasts raw-CSI scaling with aggregation and compression results.
Limitations and future work: Channel-view usefulness is setting-dependent: amplitude dominates EHealth, while calibrated phase helps selectively in WiMANS.Future systems should choose channel views adaptively rather than assume that more views always help.
Limitations and future work: Moderate unenrolled-person detection makes Argus better suited to shortlisting likely identities than final identification.The limitation concerns open-set rejection of unknown users.
Governance: Responsible deployment requires informed consent, visible opt-outs, retention limits, and local processing, and Argus should not serve as a covert identity gate.Unknown users can still receive confident closed-set identities without a separate verification layer.

Conclusion

Argus combines compact statgram inputs with a lightweight decoder-only Transformer for passive Wi-Fi CSI identification. It achieves strong aggregated accuracy on a 154-subject split and remains close to leading WiMANS baselines with substantially lower inference cost.

Conclusion: 84.85% ± 1.31% Top-1 and 99.26% Top-5 accuracy follow 60-second aggregation on a 154-subject split.These results use Argus’s compact statgram inputs and lightweight decoder-only Transformer.
Conclusion: 1.23 points is Argus’s average gap to the strongest WiMANS baselines.Argus achieves this while using 27× fewer inference FLOPs.
Conclusion: Compact CSI statistics support scalable passive identification, while robust deployment still requires stronger open-set capabilities.The passage identifies deployment robustness as an unresolved requirement.

Hardware Specifications

Experiments ran on an Apple Mac Studio with an M4 Max SoC, using PyTorch with Apple Metal Performance Shaders rather than CUDA.

Hardware Specifications: Experiments used an Apple Mac Studio with an M4 Max SoC, 16-core CPU, 40-core integrated GPU, and 128 GB unified memory.The CPU comprised 12 performance cores and 4 efficiency cores.
Hardware Specifications: The software stack included Python 3.12.11 and PyTorch 2.7.1 with the Apple Metal Performance Shaders backend enabled.CUDA was not used because the experiments ran on Apple Silicon.

Additional EHealth Details

EHealth uses a controlled 5 GHz CSI setup with 234 usable subcarriers and a standardized statgram construction. The section also specifies robust normalization and deterministic procedures for measuring runtime and compute.

EHealth collection and training: EHealth uses a 5 GHz router, laptop client, and single-antenna Raspberry Pi 4B probe, positioned 1 m from the participant.CSI was extracted across 256 subcarriers, with 234 usable; Argus training uses AdamW with the listed regularization and optimization settings.
Statistic construction: 20 statistic rows include temporal real-FFT power bands, with DC removal and equal low, middle, and high frequency splits.Each view is normalized per window using the median and interquartile range, floored at ϵ = 10−4.
Dataset-specific statgrams: 3×20×64 is the EHealth statgram shape, formed by averaging 234 usable subcarriers into W = 64 grouped bins per view.WiMANS instead uses its 30 Intel 5300 subcarriers directly, producing an 18 × 20 × 30 statgram from nine links.
Runtime and compute accounting: 30 warm-up passes and 120 timed repeats define the runtime measurements, with accelerator synchronization when available.FLOPs are computed from architecture-level multiply-add counts across patch embedding, attention, feed-forward layers, and classification.
Runtime and compute accounting: Segment-level compute multiplies single-window cost by the number of aggregated windows, using deterministic estimates rather than wall-clock latency.Wall-clock measurements may vary with kernel dispatch and background system load.

Additional WiMANS Details

The final WiMANS configuration adapts the EHealth Transformer backbone to a six-way multi-label task using an 18 × 20 × 30 statgram and 4 × 10 patches. Diagnostics examine adaptation, token granularity, model capacity, input compression, and cross-environment adaptation.

Final configuration: The final WiMANS model uses an 18 × 20 × 30 statgram, 4 × 10 patches, and a six-way multi-label head trained with weighted binary cross-entropy.It retains the same Transformer backbone as EHealth.
Statistic rows: EHealth statgram diagnostics distinguish amplitude, calibrated phase, and phase-delta statistic rows, zero-filling reserved or remaining rows as specified.Amplitude and calibrated phase use rows 1–18, while phase-delta uses rows 1–8.
Diagnostics: WiMANS diagnostics include patch-token granularity versus accuracy on classroom 2.4 GHz and input compression on the final model.The diagnostic set also covers the adaptation ladder, model capacity, and cross-environment adaptation.