Researchers have developed a groundbreaking artificial intelligence system called WhoFi that can identify and track individuals through walls using only Wi-Fi signals, achieving an impressive 95.5% accuracy rate.
This innovative approach represents a significant leap forward in biometric identification technology, offering a privacy-preserving alternative to traditional camera-based surveillance systems.
The system, detailed in recent research, leverages Channel State Information (CSI) extracted from standard Wi-Fi transmissions to create unique biometric signatures for each person.
Unlike conventional surveillance methods that rely on visual data and struggle with poor lighting, occlusions, and camera angles, WhoFi operates entirely through radio frequency analysis, enabling it to “see” through walls and other obstacles.
How Wi-Fi Becomes a Biometric Scanner
The technology works by analyzing how Wi-Fi signals interact with human bodies as they move through an environment.
When electromagnetic waves from Wi-Fi routers encounter a person, the signals are altered in unique ways based on individual physical characteristics, including internal structures like bones and organs that visual systems cannot detect.
“Wi-Fi signals offer several advantages over camera-based approaches: they are not affected by illumination, they can penetrate walls and occlusions, and most importantly, they offer a privacy-preserving mechanism for sensing,” the researchers explain.
The system captures these signal distortions in the form of CSI data, which provides detailed measurements across multiple antennas and frequencies.
The WhoFi pipeline processes this CSI data through a sophisticated deep neural network featuring a Transformer-based encoder architecture.
The system was trained using an “in-batch negative loss” function, which enables it to learn robust and generalizable biometric signatures by maximizing similarities between samples from the same person while minimizing similarities between different individuals.
Superior Performance Across Multiple Metrics
Testing on the publicly available NTU-Fi dataset, which includes data from 14 subjects performing various activities, WhoFi demonstrated exceptional performance.
The Transformer-based model achieved not only the 95.5% Rank-1 accuracy but also scored 98.1% for Rank-3 accuracy and an impressive 88.4% mean Average Precision (mAP).
The research team compared three different neural network architectures: LSTM, Bi-LSTM, and Transformer models, with the Transformer consistently outperforming the others.
This success stems from the Transformer’s self-attention mechanism, which excels at capturing long-range temporal patterns in Wi-Fi signal sequences that are crucial for accurate person identification.
The system addresses critical limitations of existing surveillance technology while maintaining privacy. No visual data is collected or stored.
This breakthrough could revolutionize security systems, intelligent building management, and healthcare monitoring applications, offering robust human identification capabilities that work regardless of lighting conditions, clothing changes, or physical obstructions.
