IT Home On October 19, a team of AI researchers at the Swedish Defense Research Agency (FOI) developed an AI system that can identify hidden pistols using Wi‑Fi signals. It can serve as a non-invasive alternative to visible light cameras and heavy metal detectors. The related paper was published on the arXiv platform on October 13. The system, called WiFireNet, identifies by analyzing the channel state information (CSI) generated by the disturbance caused by the environment’s Wi‑Fi signals on the objects. The entire system consists of a Wi‑Fi transmitter, a receiver, and the WiFireNet artificial intelligence model.
When a regular Wi‑Fi signal encounters an object during its transmission, it will penetrate and reflect depending on the shape, orientation, and material of the object, resulting in unique signal characteristics. This type of data is called channel state information, and analyzing it can achieve high-precision object recognition. Researchers in the field of security have previously disclosed a technique: regular Wi‑Fi routers can recognize the human body with an accuracy of 99.5%.
To build the WiFireNet system, researchers first created a dataset called WiFire Dataset. They recorded RGB videos simultaneously, and also collected the Wi-Fi signals reflected by people walking with various items (from phones to pistols). The researchers trained a deep learning model on this massive dataset, and using the video clips, they established a mapping between the visual features of the hidden pistols and their corresponding Wi-Fi features (channel state information). To ensure the system can distinguish pistols from everyday items with similar appearance and material, the training materials include records of collected plastic pistols, metal water cups, real metal pistols, and mobile phones.
After training, WiFireNet can detect hidden pistols solely based on Wi-Fi channel status information, but the accuracy varies. In a controlled experimental environment, in a small indoor space, up to four moving people can be tracked simultaneously, and the average accuracy of this system for detecting pistols is 72.5%. However, if more stringent positioning requirements are imposed, such as accurately locating the specific position of the weapon on the carrier, the accuracy drops to 53.8%. When dealing with an unfamiliar, untested indoor layout, the neural network’s recognition performance also deteriorates significantly.
The system is currently in the experimental stage, and the research team has planned the next optimization path. The prototype uses the modified old 1999 IEEE 802.11a wireless standard to verify the principles. In the next phase, modern wireless networks such as Wi‑Fi 7 and Wi‑Fi 8 will be adopted to obtain higher-resolution channel status information and improve detection performance. The research team will open the project to the public by releasing the code, models, and datasets, thereby promoting further research in this field.
IT Home has noticed that monitoring technologies based on Wi‑Fi perception are receiving increasing attention, and they do not require complex equipment or deployment. However, critics believe that this technology can be widely deployed, and the lack of a corresponding regulatory framework could lead to serious privacy risks.