Cross-Domain Inference for Human Localization: Applying Wi-Fi RSSI Data to CSI-Trained Models
Ariel Duschanek-Myers, Thomas Welsh, Helmut Neukirchen
cs.CR · cs.LG · cs.NI
Quick verdict
Feeds RSSI into a CSI-trained pose model, reporting ~0.80 confidence when occupied — but confidence is never checked against real positions, in one session.
Summary
This paper tests whether a model trained on Wi-Fi Channel State Information (CSI) — the Person-in-WiFi 3D pose estimator (Yan et al., CVPR 2024) — can localize people when fed Received Signal Strength Indicator (RSSI) values, which commodity IoT devices can read without elevated OS permissions. In a small campus room replicating the original training layout (three ESP32-C3 transmitters, a Raspberry Pi 4 receiver), the authors recorded ~2 minutes of empty-room baseline and a roughly four-minute occupied session with nine annotated movement events, synchronized to webcam video. After converting dBm to linear amplitudes, subtracting the empty-room baseline, and applying a tuned pipeline (noise gate 0.00095, sigmoid K=4000, angular embedding mimicking CSI phase), the unmodified model's per-frame maximum confidence reached 0.9217 (0.7189 without embedding), with roughly 0.80 average confidence while the room was occupied; torso keypoints showed the highest prediction spread. The main limitation is evaluative: the paper reports model confidence, never measuring predicted positions against the collected video ground truth, so localization accuracy is unestablished. Additional concerns include a single participant and session, preprocessing hyperparameters apparently selected on the evaluation recording, empty-room confidence reaching 0.6529, and an unspecified RSSI-to-CSI-tensor mapping. The privacy threat model (SDK-accessible RSSI enabling surveillance) is plausible, but the evidence supports presence/movement detection rather than validated localization.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- RSSI data, without retraining, can drive the unmodified CSI-trained Person-in-WiFi 3D model through cross-domain inference to produce human pose/location predictions.
- With angular-embedding preprocessing, occupied-room predictions reach a maximum confidence of 0.9217 and roughly 0.80 average confidence, versus 0.7189 maximum without embedding.
- Torso keypoints (6, 8, 10) exhibit the highest prediction standard deviation, consistent with body mass driving signal change, and confidence statistics separate occupied from empty-room states.
- Because RSSI is readable by ordinary applications without elevated OS permissions, a wide range of commodity IoT devices could enable location privacy invasion in Wi-Fi-dense environments.
Strengths
Practical, well-motivated threat model
Targets RSSI as SDK-readable on commodity devices (Android smart TV example, Sec. I) versus CSI's elevated OS permissions and specialized drivers (Sec. II) — a genuinely underexamined surveillance vector for IoT-dense environments.
Disciplined signal hygiene and controls
Empty-room baseline recorded before every evaluation, diagnosis of power-adapter/shielded-cable noise on ESP32_AP3 (Sec. III.C), WPA supplicant disabled, laptop Wi-Fi turned off during collection — careful experimental hygiene for RF work.
Systematic preprocessing ablation
Transparent, code-level preprocessing
Candid threats-to-validity section
Weaknesses
Confidence is conflated with localization accuracy
Sample size of one
Hyperparameters tuned on the test recording
Weak and self-contradictory empty-vs-occupied separation
Unspecified input mapping and a math slip
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.17204, University of Iceland, IEEE copyright notice) asks a practical privacy question: can the RSSI value that ordinary applications read from commodity Wi-Fi devices — without the elevated OS permissions and specialized drivers CSI requires — be fed into an existing CSI-trained sensing model for human localization? Using the Person-in-WiFi 3D pose estimator (Yan et al., CVPR 2024) unmodified, the authors report that cross-domain inference works: per-frame maximum confidence of 0.9217 with angular embedding (0.7189 without) and roughly 0.80 average confidence while a person moves in the room (Sec. IV.B, Figs. 7-9). The abstract escalates this to 'predict locations with approximately 80% confidence' and to wide-scale IoT privacy invasion; Sec. IV.B additionally claims the model found RSSI 'generally as meaningful as CSI' — a claim with no CSI control condition behind it. Both statements outrun the evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can Wi-Fi RSSI really be used to track people without special hardware or permissions?
This paper offers a first indication: an existing pose model trained on CSI (Person-in-WiFi 3D) produced high-confidence predictions (per-frame maximum 0.9217; roughly 0.80 average while a person moved) when fed preprocessed RSSI from three ESP32-C3 transmitters and a Raspberry Pi. However, the study measures model confidence, not actual positional accuracy against ground truth, and used a single person in a single session — so it demonstrates presence/movement-detection potential, not validated tracking.
What equipment was used to collect the RSSI data?
A roughly 3.5×4 m campus room with three ESP32-C3 boards acting as Wi-Fi access points and a Raspberry Pi 4 as the receiver, running a shell-script scanner under Raspbian Linux; a USB webcam recorded the participant for ground truth. The layout deliberately mirrored the training environment of the Person-in-WiFi 3D project.
Does this mean smart TVs or IoT devices can spy on my location today?
Not demonstrated. The paper argues RSSI is readable by ordinary apps without special drivers, making the privacy threat plausible in Wi-Fi-dense environments, but the evidence is one controlled lab session with one participant. The authors themselves note predictions may be internally consistent yet 'inaccurate compared to reality', and no positional error was ever measured.
Is the paper's code or dataset publicly available?
No release of the authors' own RSSI dataset or full pipeline is stated. The paper reuses the public Person-in-WiFi 3D model and training data and includes two short preprocessing code listings (linear normalization and angular embedding), but the input-assembly code and collected data are not stated as available.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-9 cannot be inspected from extracted text; the empty-vs-occupied separation judgment rests on the authors' figure descriptions, which contain an internal contradiction around Fig. 7.
- Text extraction shows OCR artifacts (e.g., the linearization formula appears as '10 RSSI 2×10', and the Fig. 1 caption interleaves with body text); numbers were cross-checked against prose, but minor misreads are possible.
- Only the Listing 1-2 snippets are embedded; the RSSI-to-model tensor-assembly code is absent, so the exact input construction could not be verified.
- The IEEE copyright block and footnotes interleave with front matter; section boundaries and some table column labels (Table I appears to have five headers but four value columns) are approximate.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.