Swedish researchers have trained an AI to spot concealed handguns by reading how bodies and belongings distort Wi-Fi signals, and in controlled tests it reached 72.5% average precision.
The team at the Swedish Defense Research Agency (FOI) described the system, called WiFireNet, in a paper posted to arXiv on October 13. It needs a Wi-Fi transmitter, a receiver and the model. Signals crossing a room bounce and bend depending on an object's shape, orientation and material. Those measurements are known as channel state information (CSI), and WiFireNet reads them to guess what a person is carrying.
To train it, the researchers recorded video and reflected Wi-Fi signals at the same time while people walked around with different objects, then used the footage to label the signals. The props included a real metal handgun, a plastic handgun, a metal water bottle and a phone, picked to test whether the system could tell a gun from lookalikes in shape or material.
Precision, the share of the system's handgun alerts that were correct, averaged 72.5% with up to four people moving in a small indoor space. Asking it to also pinpoint where on the person the weapon sat dropped that to 53.8%. The model also struggles in indoor layouts it has not seen before.
That last weakness is the one to watch. A detector that needs retraining for every room is a lab result, not a replacement for a metal detector, which at least works the same way anywhere. The prototype also ran on a modified version of the 1999 802.11a standard. The researchers said later phases will use Wi-Fi 7 and Wi-Fi 8, whose higher-resolution CSI should help. They released the code, models and dataset as open source.
The privacy question travels with the technique. Security researchers previously showed that ordinary Wi-Fi routers can identify people with 99.5% accuracy, and critics say the spread of Wi-Fi sensing has outrun any regulatory framework. If gun detection gets good enough to deploy, the open question is who decides where the same signals may be read.