Trust-Centric Mitigation of Time-Warping Adversarial Attacks in Wearable Sensor Platforms
Abstract
Wearable sensors are increasingly deployed in healthcare, fitness, and human activity recognition (HAR). Adversarial attacks on deep learning models used in HRA increase security concerns in consumer goods. Adversarial time-series perturbations decrease the accuracy of prediction models. This paper investigates the impact of Time-Warping Adversarial Attacks (TWAA), which manipulate the temporal axis of sensor signals while preserving amplitude ranges. Thereby, generating adversarially misleading samples. We propose a trust-centric framework comprising attack execution, detection, and mitigation modules, and evaluate it on two benchmark datasets (UCI HAR, WISDM and MHEALTH) using baseline deep models (CNN, LSTM, and CNN–LSTM). Experimental results show that TWAA significantly degrades classification performance. It reduces the adversarial accuracy by up to 30% and achieving attack success rates exceeding 40% at moderate perturbation parameters. Our detection mechanism achieves AUROC > 0.85 across different types of TWAA, while the mitigation strategy restores model accuracy to above 80% with inference latency under 0.01 ms/sample. These findings demonstrate both the severity of temporal adversarial threats in HAR and the effectiveness of our proposed trust-centric defense. Unlike existing mechanisms that rely on adversarial training or offline robustness enhancement, the proposed framework enables runtime detection of physically plausible time-warping adversarial attacks and applies conditional input-level temporal repair using dynamic time warping. Thus, providing a holistic pipeline for the analysis, detection and mitigation of TWAA on wearable consumer goods.