A Wi-Fi network may soon do more than connect phones and stream video. Researchers at Sweden’s Defence Research Agency, or FOI, have developed WiFireNet, an experimental AI system that can detect concealed handguns by studying how wireless signals change as they pass around people and objects. The work was detailed in an October 13 arXiv paper by Erik Persson, David Gustafsson, and Tomas Wilkinson. Tom’s Hardware first reported on the project’s weapon-detection potential.
The system relies on channel state information, or CSI: measurements describing how a radio signal is altered by its surroundings. A handgun’s shape, position, and materials affect the signal differently from a phone, water bottle, or similarly shaped plastic object. WiFireNet uses a transmitter, a receiver, and a neural-network pipeline to turn those distortions into an object-detection result. The model can identify and localize targets from CSI alone once it has been trained; camera imagery is used during training as a source of annotations and visual supervision, not during inference.
To train the system, the FOI team built the WiFire dataset by synchronizing RGB footage with wireless measurements as between zero and four people moved indoors carrying different objects. The collection included a deactivated metal handgun, a plastic handgun, a metal water bottle, and a mobile phone, giving the model examples of both the target and ordinary items that could confuse it. Data came from two rooms measuring roughly 8 and 12 square meters, with the smaller space producing stronger multipath reflections that generally made detection easier. The researchers have released the WiFire dataset along with WiFireNet’s code and models.
The headline accuracy figure needs careful unpacking. In the reported large-room experiment, WiFireNet achieved 72.5% average precision for the handgun class at a 25% intersection-over-union threshold, but that fell to 53.8% when the system had to place the weapon more precisely at a stricter 50% threshold. Overall detection results were lower when the model had to classify and localize multiple object categories. Those numbers show that wireless sensing can carry useful information about a concealed weapon, but they are not yet evidence of a dependable security checkpoint.
There are also important environmental caveats. The dataset was collected in only two relatively clean indoor rooms, without the clutter and unpredictable movement found in airports, schools, stadiums, or busy shops. The paper does not establish performance in other buildings, outdoors, or previously unseen environments, and its sample-level data split leaves open questions about how well the model would handle new people and sessions. WiFireNet currently uses a modified implementation of the older IEEE 802.11a standard at 5.68 GHz and 20 MHz bandwidth, rather than a production Wi-Fi security system.
That makes the project more of a research platform than an imminent replacement for cameras or metal detectors. Conventional cameras can miss weapons hidden by clothing, while metal detectors usually require people to pass through controlled entry points and cannot readily distinguish a handgun from other metal objects. Wi-Fi sensing could, in theory, monitor a space without requiring a person to stop or stand in front of a camera. But its unobtrusive nature also raises a harder question: whether increasingly common wireless infrastructure should be allowed to perform surveillance without people realizing it.
For now, the FOI researchers frame WiFireNet as a possible complement to existing systems, not a standalone defense layer. Releasing the dataset, code, and trained models gives other researchers a way to test the approach in messier environments and improve its generalization. The next meaningful milestone will be seeing whether newer Wi-Fi hardware and richer CSI measurements can turn a promising laboratory demonstration into a system that remains accurate across different rooms, crowds, clothing, and everyday wireless interference.
