The Pulse
FOI Researchers Train WiFireNet to Detect Handguns Through Wi-Fi
Researchers at Sweden’s FOI developed WiFireNet, a model that uses Wi-Fi signal measurements to detect concealed handguns in tests involving up to four people. The study reports different results depending on how closely predicted locations
AI.info Team · Written with AI ·
“Our baseline results demonstrate that WiFi signals indeed contain sufficient information to detect concealed handguns carried by humans in groups of up to four people and in the presence of other commonly carried items.”
The authors, Erik Persson, David Gustafsson and Tomas Wilkinson, write in “Concealed Handgun Detection using WiFi Signals.”
The three researchers at the Swedish Defence Research Agency, FOI, published their WiFireNet paper on arXiv on October 13, 2026. The system analyzes channel state information (CSI), measurements of how a wireless signal changes as it travels between a transmitter and a receiver. Rather than search a camera image, the model uses those changes to estimate where people and carried objects appear.
The authors describe Wi-Fi sensing as a possible supplement to cameras and metal detectors, not a replacement for either. Cameras can miss occluded objects, while conventional metal detectors require people to pass through a checkpoint. WiFireNet explores whether radio measurements can help spot a concealed handgun in a room where several people are moving.
How WiFireNet learns from camera footage
For training and evaluation, the team built the WiFire dataset, pairing Wi-Fi measurements with RGB video recorded at the same time. The video supplied labels to teach the model what to look for; during inference, WiFireNet uses CSI alone. That distinction matters: cameras helped prepare the model, but the reported detection step does not require a live camera image.
The dataset includes four object classes: a metal handgun, a plastic handgun, a metal water bottle and a mobile phone. The decoy objects test whether the model can distinguish a real handgun from items that share its shape or material. The study used a concealed, deactivated handgun, and included scenes with up to four people carrying common objects.
Researchers collected data in two rooms: an 8-square-meter room with up to two people and a 12-square-meter room with up to four. Their wireless setup used a modified, open-source implementation of IEEE 802.11a, operating at 5.68 GHz with 20 MHz bandwidth. Alongside standard 802.11a measurements, they tested a “WiFi-plus” configuration that used more pilot subcarriers to provide additional CSI information.
The scores depend on what counts as a match
WiFireNet does not produce one simple accuracy figure. The paper reports recall and mean average precision (mAP) at two intersection-over-union thresholds, which measure how closely a predicted box overlaps the labeled location. In the large-room test using standard Wi-Fi, overall recall was 91.5% at the 25% overlap threshold, while overall mAP was 50.3% at that same threshold.
The often-cited 72.5% and 53.8% figures refer to average precision for the handgun class specifically, under the standard Wi-Fi configuration: 72.5% at the 25% overlap threshold and 53.8% at 50%. They are not the system’s overall precision or a finding that it correctly identifies that share of all people carrying guns. The narrower threshold requires a closer match between the predicted and labeled box, and the handgun-class score falls under that stricter test.
More CSI data did not automatically improve results. In the paper’s baseline table, the standard configuration outperformed WiFi-plus on overall mAP in the large room: 50.3% versus 39.9% at the 25% threshold. The authors suggest the cleaner test environment may have limited the benefit of extra signal information, and say more complex settings need further study.
Two rooms do not establish checkpoint readiness
The experiment is a proof of concept, not a field trial across varied buildings or crowds. Data came from two indoor rooms, and the researchers split the dataset into training and test sets at the level of individual CSI samples. The paper does not report validation across a broad range of unseen layouts, which leaves open how well the model would transfer to a different venue.
FOI’s work provides a dataset, model and code for further research, while its reported results show both the promise and limits of this particular setup. The strongest handgun-class result is tied to a lenient overlap threshold; performance is lower under the stricter one. For now, WiFireNet shows that Wi-Fi measurements can carry useful information about concealed handguns in controlled tests, not that ordinary wireless networks can reliably screen people in everyday settings.