正文 · AI 翻译
IT Home reported on October 19 that a group of AI researchers at the Swedish Defence Research Agency (FOI) has developed an AI system that can use Wi‑Fi signals to identify concealed handguns, offering a noninvasive alternative to visible-light cameras and bulky metal detectors. A related paper was published on arXiv on October 13. The system, called WiFireNet, identifies objects by analyzing channel state information (CSI) generated when objects disrupt ambient Wi‑Fi signals. The complete system consists of a Wi‑Fi transmitter, a receiver, and the WiFireNet AI model.
When ordinary Wi‑Fi signals encounter objects as they travel, they pass through or reflect off them depending on the objects’ shape, orientation, and material, producing distinctive signal characteristics. These data are known as channel state information, and analyzing them can enable object recognition with relatively high accuracy. Security researchers have previously disclosed a technique in which an ordinary Wi‑Fi router can identify people with an accuracy of up to 99.5%.
To build WiFireNet, the researchers first created a dataset called the WiFire Dataset. They recorded RGB video while collecting Wi‑Fi signals reflected by people walking while carrying various items, from mobile phones to handguns. The researchers trained a deep learning model on this large dataset, using the video clips to map the visual features of concealed handguns to their corresponding Wi‑Fi features (channel state information). To help the system distinguish handguns from everyday objects with similar shapes and materials, the training data included recordings of plastic handguns, metal water cups, real metal handguns, and mobile phones.
After training, WiFireNet can detect concealed handguns using only Wi‑Fi channel state information, though its performance varies. In a controlled experiment in a small indoor space, with up to four moving people tracked simultaneously, the system achieved an average precision of 72.5% for handgun detection. But under stricter localization requirements, such as pinpointing a weapon’s exact position on the person carrying it, precision fell to 53.8%. The neural network’s recognition performance also deteriorated markedly in unfamiliar indoor layouts that had not been tested.
The system remains experimental, and the research team has outlined plans for further improvements. The prototype uses a modified version of the old 1999 IEEE 802.11a wireless standard to validate the concept. The next stage will use modern Wi‑Fi 7 and Wi‑Fi 8 networks to obtain higher-resolution channel state information and improve detection performance. The research team has made the project open source, releasing its code, model, and dataset to support further research in the field.
IT Home noted that Wi‑Fi sensing-based surveillance technology is attracting increasing attention because it does not require complex equipment or deployment. Critics, however, say that the technology’s potential for widespread deployment, coupled with a lack of a corresponding regulatory framework, poses serious privacy risks.