<p>Smartwatches facilitate the efficient and continuous collection of high-resolution health data through integrated sensors, supporting various health-related applications such as sleep monitoring, fall detection, and stress detection. While these advancements offer significant benefits for health tracking and research, they also raise serious privacy concerns, as health data often contains sensitive personal information. This study presents a modular attack framework consisting of four similarity-based re-identification attacks that exploit vulnerabilities in de-identified time-series health data. Using Dynamic Time Warping (DTW) to measure similarity between samples, we evaluate our approach on the WESAD dataset (15 participants) and two synthetic datasets generated via Generative Adversarial Networks (GANs), each comprising up to 1000 subjects. Despite the use of privacy-preserving measures, our attacks achieved 100% re-identification accuracy on the WESAD dataset and over 93% on the synthetic datasets, revealing that even short sequences of sensor data can compromise user anonymity. To address this issue, we explore additional privacy protection through noise injection and conduct a case study assessing the impact of varying noise levels on both re-identification risk and data utility for a stress detection task. The results show that adding noise significantly reduces the re-identification risk while still maintaining strong classification performance, emphasizing the need for advanced privacy models beyond traditional de-identification to ensure the safe use of health data in wearable technologies.</p>