Reproducible benchmark of wavelet-enhanced intrabody communication biometric identification

Area of Science:

• Biometrics
• Signal Processing
• Wearable Technology

Background:

• Intrabody communication (IBC) channels offer unique physiological data for wearable biometric identification.
• Previous high accuracy claims in IBC biometrics are often due to data leakage, inflating performance metrics.

Purpose of the Study:

• To establish a public, leakage-free benchmark for IBC biometrics using a 30-subject dataset.
• To systematically compare different signal representations (frequency-domain, time-frequency) for biometric identification.
• To evaluate the performance and efficiency of feature extraction methods on embedded systems.

Main Methods:

• Developed a leakage-free benchmark with subject-wise 80/20 data splits, repeated five times.
• Compared Scattering, Discrete Wavelet Transform (DWT) statistics, and spectral representations with machine learning classifiers (LightGBM, Random Forest, MLP, SpectralCNN).
• Assessed embedded performance on an STM32F446RE microcontroller for latency and energy consumption.

Main Results:

• The strongest classical configuration (Scattering + LightGBM) achieved 54.0% accuracy under the leakage-free benchmark.
• Wavelet features (db4-DWT, lifting-based) with Random Forest outperformed the baseline, reaching 51.6% and 49.3% accuracy, respectively.
• Exploratory neural network analyses showed higher accuracy (up to 83.7%) but are not leakage-free; confusion analysis suggests a potential biometric ceiling.

Conclusions:

• The established benchmark provides a reliable evaluation of IBC biometric systems, highlighting current accuracy limitations.
• Lifting-based wavelet features offer a promising balance of accuracy, low latency, and low energy consumption for embedded biometric applications.
• Open-sourcing code and data facilitates reproducibility and future research in robust IBC biometric identification.