Features of technical implementation of biometric authentication system based on Wi-Fi channel state information analysis

This paper addresses the problem of contactless biometric authentication for industrial enterprise access control, where traditional methods – fingerprint scanners, face recognition terminals, and proximity cards – face practical limitations due to hygiene requirements, personal protective equipment, and high personnel throughput at checkpoints. The proposed solution is based on Wi-Fi Channel State Information (CSI) analysis combined with machine learning, enabling passive identification of employees as they pass through a designated control zone without any deliberate interaction with the system.

The physical basis of the approach lies in the unique way each individual modulates radio signal propagation. A person's anthropometric characteristics determine the effective scattering cross-section of their body, producing distinctive amplitude patterns across signal subcarriers, while involuntary micro-movements such as breathing and cardiac activity introduce low-frequency phase modulations in the 0.1…3 Hz range. These individual-specific patterns constitute the biometric signature extracted by the system. Since manually engineering features that capture such subtle radio-physical phenomena is practically infeasible, a machine learning model is essential for automatic discovery of discriminative representations in high-dimensional time-series CSI data.

The system hardware employs ESP32 microcontrollers placed 3 meters apart at 1.5-meter height, forming a control zone in which human presence substantially modulates the received signal across 52 OFDM subcarriers sampled at 100 Hz. The signal preprocessing pipeline consists of four sequential stages: temporal averaging to suppress high-frequency noise, a robust Hampel outlier filter based on median absolute deviation, a fifth-order Butterworth low-pass filter isolating the biomechanically relevant frequency band below 10 Hz, and linear phase calibration compensating for hardware clock instabilities. The processed data is segmented into one-second overlapping windows, each forming a classification sample.

The core contribution of the work is a specialized Dual-Branch Transformer architecture that processes amplitude and phase CSI characteristics through independent encoder branches. This design choice is motivated by the physically distinct nature of the two signal components: amplitude primarily encodes anthropometric information, while phase is more sensitive to fine-grained displacement dynamics. Processing them jointly would risk one modality dominating the learned representation. Each branch applies a linear projection of the 52-subcarrier input to a 32-dimensional space, adds sinusoidal positional encoding to preserve temporal structure, and passes the sequence through a Transformer encoder with four attention heads. The self-attention mechanism explicitly models pairwise dependencies across all time steps within the window, making it well-suited for capturing periodic micro-movement patterns that span several hundred milliseconds. Branch outputs are mean-pooled, concatenated, and passed to a linear classifier. The resulting model contains only 15,742 parameters, making it trainable on modest enterprise-scale datasets.

Experimental evaluation on a dataset of three employees and an Empty Zone class achieved an overall classification accuracy of 92.3%, with the Empty Zone class reaching perfect 100% accuracy – a critical property ensuring the system never grants access when no person is present. All misclassifications occurred exclusively between employee classes, with per-user accuracy ranging from 81.0% to 92.2% depending on anthropometric distinctiveness. An additional experiment demonstrated that raising device placement height from 1.0 to 1.5 meters improved overall accuracy to 96.0%, due to more favorable signal incidence angles and reduced floor-reflection interference.

The real-time recognition module operates as a five-state finite automaton, requiring three consecutive presence detections for initial triggering, five consistent predictions for signal stabilization, and a five-second majority voting procedure for final identification. End-to-end latency from zone entry to identification onset is under one second. Field tests confirmed recognition confidence levels of 85.7…96.5% across successful sessions, with the system running on a standard personal computer without hardware acceleration.

The results demonstrate the practical feasibility of passive Wi-Fi CSI-based biometric authentication for industrial access control, establishing a foundation for deployments in environments where contact-based biometrics are constrained by hygiene, safety, or operational requirements.