A Differential Privacy Approach for Privacy-Preserving Multi-Modal Stress Detection

There has been a noticeable spark in the interest towards leveraging physiological data in various Internet of Medical Things (IoMT) applications. However, due to the sensitive nature of physiological time-series data, there is an urgent need to maintain the privacy of these data when used for training machine learning (ML) models. While these models show benefits for detecting various medical conditions from the time-series data (e.g., stress detection), the risk of undermining the privacy of users’ data outweighs the sheer benefits attained from these models. Via this work, we aim to primarily focus on studying differential privacy as a viable privacy-preserving method for training ML models. In particular, we focus on classification tasks involving multiple data streams and utilize a differential private optimizer. We present a detailed analysis involving single and multimodal streams and empirically evaluate the role and impact of differential privacy on stress detection using various data streams. This preliminary work sets the foot-steps towards exploring the trade-off between data utility and privacy in the IoMT domain with application to stress detection.