On-Device ECG Personal Authentication via Autocorrelation Profile and Selective Precision Binary Neural Network

This paper proposes a lightweight personal authentication system that combines autocorrelation profile (ACP)-based feature representation with a selective precision binary neural network (SPBNN) for practical ECG-based authentication in on-device AI environments such as massage chairs. The ACP-based feature representation reduces sensitivity to motion artifacts and signal instability in dry-electrode environments, while a selective LSB activation mechanism based on MSB/LSB decomposition and a learnable threshold Δ partially recovers the authentication performance lost through binary quantization. The primary quantitative evaluation was conducted using the public HMC dataset comprising 154 subjects. In addition, a self-collected pilot dataset of 10 subjects was used to examine the preliminary feasibility of the proposed method under simulated massage-chair operating conditions. In the fixed-split evaluation, the proposed model achieved F1-scores exceeding 90% on both datasets (90.74% for 154 subjects and 90.13% for 10 subjects), with model sizes of 0.33 MB and 0.12 MB, respectively. Furthermore, the CNN-relative performance retention rates were 93.74% and 93.87%, compared with 86.84% and 86.86% for the pure BNN. These results demonstrate that the proposed SPBNN achieves a favorable balance between model compactness and authentication performance, while the pilot evaluation suggests its potential applicability in resource-constrained on-device AI environments.