Trustworthy Stress Detection in Consumer Wearables: Poisoning-Resilient Learning with Feature-Gated DualNet and Generative Restoration

Wearable devices are increasingly used for continuous stress monitoring, yet their reliability under adversarial data manipulation remains largely unexamined. This paper presents the first systematic study of training-time poisoning in consumer-grade stress detection, where adversaries perturb or inject malicious samples during model training. We frame stress recognition as a transaction-level classification problem over short, multimodal physiological windows, explicitly accounting for class imbalance, noise, and contamination. Across two widely used benchmarks (WESAD and DEAP), we show that classical machine learning models (Logistic Regression, SVM, Random Forest) exhibit severe degradation under poisoning, with Macro-F1 dropping by up to 30% and stress recall collapsing below 0.50 at 40% contamination. To address this, we introduce Feature-Gated DualNet (FG-DualNet), a dual-branch neural architecture that combines adaptive feature gating with context-preserving representation learning. We further propose a generative restoration defense that leverages outlier filtering and CTGAN-based synthesis to reconstruct poisoned minority classes. This proactive defense operates at training time by identifying anomalous transactions and reconstructing stress-class distributions before model retraining. Together, these techniques sustain up to 20–25 percentage points higher recall under severe poisoning while maintaining interpretable modality-level importance (e.g., electrodermal activity and respiration). Our findings establish poisoning robustness and physiological interpretability as critical design objectives for trustworthy, stress-aware consumer electronics, and provide a reproducible pipeline for evaluating resilience in future wearable sensing systems.