A new method, called WhoFi, enables the re-identification of individuals based on Wi-Fi signals, without using visual data.
Italian researchers developed an alternative to camera surveillance that uses Wi-Fi signals to identify individuals. The technique utilizes Channel State Information (CSI) to extract biometric features from radio signals. The findings were compiled in a paper titled “WhoFi: deep re-identification of individuals via Wi-Fi channel signal encoding”. The research was published in Arxiv.
Alternative to Camera Surveillance
Researchers propose a method to recognize individuals using Wi-Fi signals instead of camera images. The technique, called WhoFi, analyzes Channel State Information (CSI) to extract biometric features from radio signals. This information is then processed by a modular deep learning model, which includes Transformer networks. According to the researchers, the approach offers a robust alternative to traditional camera-based systems that face issues such as poor lighting, visual obstruction, or unfavorable camera angles.
In the method, a person’s physical characteristics—such as body shape, bone structure, or composition—affect how Wi-Fi signals propagate in a space. These signal distortions are interpreted by the system as unique biometric signatures. Unlike optical systems, Wi-Fi signals can also penetrate walls and objects and do not rely on light, enabling deployment in complex environments.
Deep Learning
The researchers trained their deep learning model with a so-called in-batch negative loss function. This technique allows for learning distinctive features without the need for manually labeled data pairs. For processing the time series with CSI data, they tested various networks, including LSTM, Bi-LSTM, and Transformer architectures. The evaluation was conducted on the public NTU-Fi dataset.
Experiments show that WhoFi achieves results comparable to state-of-the-art methods based on camera images. Additionally, the researchers examined the impact of factors such as sequence length, model complexity, and data augmentation on performance. According to the team, this research demonstrates that wireless biometrics can be a scalable and privacy-friendly alternative to video surveillance.
According to the authors, the system offers potential for applications where visual perception is unreliable or undesirable, such as in poorly lit spaces, through walls, or in privacy-sensitive environments.