Efficiently Updating ECG-Based Biometric Authentication Based on Incremental Learning

Figure 1. The basic waveform of electrocardiograms (ECG) with five fiducial points.

Figure 2. Sample states according to Karush–Kuhn–Tucker (KKT) conditions.

Figure 3. Scheme to implement and evaluate the proposed ECG-based authentication.

Figure 4. Synchronized average of ECG signals from 11 subjects. The shaded areas represent the standard deviation in the signals.

Figure 5. Synchronized average of ECG signals of the subject 10 for six days. The shaded areas represent the standard deviation in the signals.

Figure 6. Detected Q, R and S peaks of an ECG signal.

Figure 7. Detected P and T peaks of an ECG signal.

Figure 8. ECG features for the authentication: four amplitudes (a), three angles (b), and six temporal features (c) including RR interval.

Figure 9. Objective function values for hyperparameter tuning using Bayesian optimizer. The optimal SVM parameters are C = 7 (box constraint) and σ = 2 (kernel scale).

Figure 10. Updating decision boundary of incremental SVM for subject 4 data set over five-day incremental learning. The darker dots from days 2 to 5 represent the incrementally trained data.

Figure 11. The accuracy, false acceptance rate (FAR) and true acceptance rate (TAR) of 11 subjects over five days of incremental learning (a) and the average (b) tested using the data of day 6.

Figure 12. The accuracy, FAR and TAR of 11 subjects over five days of incremental learning (a) and the average (b) tested using the data recorded a day after the latest training.