WiFire - Concealed handgun detection using WiFi signals
Official dataset repository for the paper: "Concealed Handgun Detection using WiFi Signals".
Overview
Conventional surveillance systems (cameras, metal detectors) face limitations such as occlusion vulnerability, reliance on poor lighting, and spatial constraints that cause crowd congestion. WiFireNet addresses these challenges by utilizing WiFi Channel State Information (CSI) as a complementary modality capable of overcoming occlusions while operating discreetly across wide areas without restrictive spatial checkpoints.
Key Contributions
- WiFire Dataset: A novel cross-modal dataset containing aligned WiFi CSI samples and RGB video annotations featuring up to 4 simultaneous participants carrying common items (mobile phones, water bottles) and concealed threats.
- WiFireNet Architecture: A specialized deep learning model featuring a novel convolutional Modality Translation Network (MTN) that translates raw 1D CSI phase/amplitude into 2D image-space features.
- Cross-Modal Supervision: Employs a vision-based teacher model during training via a cross-modal Knowledge Distillation (KD) feature loss.
Dataset: The WiFire Dataset
The dataset contains aligned WiFi CSI bursts and RGB images, captured at 5.68 GHz with a 20 MHz bandwidth under two distinct configurations:
- WiFi-regular: Adheres to the standard IEEE 802.11a protocol, utilizing 4 pilot subcarriers with the remaining 48 linearly interpolated (~100k samples).
- WiFi-plus: Utilizes 48 out of 52 available pilot subcarriers (92.3% coverage) to provide fine-grained, radar-like resolution (~100k samples).
- Small-room setup: Captured within an 8 m² area (0-2 people) characterized by strong multi-path reflections (~40k samples).
- Large-room setup: Captured within a 12 m² area (0-4 people) with sparser reflections.
Please note: In wifi-regular/2 persons/2025-10-17 17_10_21.074780_frames there is a folder called frames_after_5000. Move the content of this folder to its parent.
This is only due to HuggingFace limiting the number of files to 10 000 per folder.
Object Classes
The dataset evaluates the network's discriminative ability across four highly similar or common classes:
Handgun(Deactivated metal handgun)Plastic Handgun(Shape distracter)Metal Water Bottle(Material property distracter)Mobile Phone(Common everyday item)
To create dataset splits and get our baseline model working, see https://github.com/persson-erik/WiFireNet.
Baseline Evaluation (mAP and Recall at 25% and 50% IoU)
| Dataset | Recall @25 | Recall @50 | mAP @25 | mAP @50 |
|---|---|---|---|---|
| WiFi-regular (Large Room) | 91.5 | 89.5 | 50.3 | 47.1 |
| WiFi-plus (Large Room) | 87.7 | 85.5 | 39.9 | 37.2 |
| Small Room (WiFi-plus) | 90.0 | 88.2 | 49.8 | 46.9 |
| WiFi-plus (<2 people) | 85.3 | 83.5 | 37.8 | 35.5 |
Collecting your own WiFi data
To collect your own data, see https://github.com/persson-erik/WiFire-data-collection
Citation
Please use the following to cite our work:
@inproceedings{persson2026concealed, author = {Persson, Erik and Gustafsson, David and Wilkinson, Tomas}, title = {Concealed Handgun Detection using WiFi Signals}, booktitle = {Proceedings of the Fifth International Conference on Pattern Recognition and Artificial Intelligence (ICPRAI2026)}, year = {2026} }