Res-PGDnet: A time frequency fusion model with physiologically constrained adversarial training for PPG-based valence–arousal state classification

Abstract

Accurate valence–arousal assessment is crucial for affective computing applications. Photoplethysmography (PPG) is a promising, practical modality for this task, but it poses challenges: signals are noise-prone, the physiological affect mapping is complex and subject-specific, and models generalize poorly. Existing methods often focus narrowly on temporal features, employ physiologically unconstrained robustness strategies, or suffer from a severe performance–efficiency trade-off. We propose Res-PGDnet to address these issues via synergistic time frequency fusion and physiologically constrained adversarial training. Its core innovations are (i) applying adversarial perturbations constrained by physiological priors to model parameters to ensure robust, plausible learning and (ii) hierarchically integrating DCT attention modules to jointly enhance temporal and spectral feature representations. On the DEAP-PPG dataset, Res-PGDnet achieves 93.47% intra-subject accuracy and 92.46% Macro-F1 score, with strong robustness (metric: 0.920) and 85.1% cross-subject accuracy. The model is efficient (12.5 M parameters, 1250 FPS), demonstrating potential for wearable-based valence–arousal classification in settings like mental health monitoring.