Robust PPG Authentication: The Role of Temporal, Individual, and State Variability

Photoplethysmography (PPG), a signal tracking blood volume changes, encodes user-specific patterns for unobtrusive authentication. This study investigates three real-world factors that affect the reliability of PPG-based authentication: temporal variability (time elapsed between training and test), user-specific variability (differences in individual identifiability), and state-dependent variability (changes in states such as stress or mood). We use the LAUREATE dataset, which includes longitudinal PPG signals and self-reported states (e.g., stress, happiness, sleep quality) from 44 participants over 13 weeks. We train a one-dimensional convolutional neural network (CNN) to identify users from 180-second PPG segments. Results show no systematic decline in performance due to temporal variability, but notable individual and daily fluctuations. Some users are consistently identified with high accuracy, while others show more variable performance, indicating identifiability varies across individuals. Most notably, the performance of the model improves when trained on data matching the test-state value, yielding up to 0.3 higher Matthews Correlation Coefficient for individual states and 0.11–0.2 improvement when states are grouped using the PANAVA-KS scale. These findings suggest that state-dependent variability, rather than time, is the primary limitation of biometric authentication systems. Robust authentication systems should incorporate multi-state training and context-aware modeling to ensure consistent performance in real-world settings.