Robust Human Information Inference against IoT Sensing Data Shift

Zhizhang Hu

Abstract: Author(s): Hu, Zhizhang | Advisor(s): Pan, Shijia | Abstract: The deployment of Internet-of-Things (IoT) systems with artificial intelligence (AI) algorithms enables a wide range of human-centered smart applications, from wearable health monitoring to personalized services through occupant identification and human activity recognition. However, when these systems are deployed in real-world environments, they often experience reduced inference accuracy due to shifts in the data. To ensure accurate inference in these dynamic settings, this dissertation explores the causes of degraded accuracy and proposes approaches to enhance system performance, minimizing the need for costly re-labeling of the new data.This work addresses data shift from three key perspectives: concept drift, dataset bias, and domain variance. Concept drift occurs when the relationship between input features and target variables changes, particularly due to individual physiological and behavioral differences. To address this, we introduce PAN Net, an adaptive model that identifies and emphasizes the most informative sensor channels, ensuring generalizable performance even in the presence of such drift. PAN Net is applied in the IOTeeth system for occlusal disease monitoring, using a harder attention mechanism to dynamically adjust the importance of each sensor channel during inference. Dataset bias arises when the distribution of classes in the training data differs from that in the deployment environment, leading to poor generalization. To combat this, we propose CIPhy, a causal inference-based framework that mitigates the impact of dataset bias by promoting the learning of invariant feature-label correlations, thereby improving model robustness.Domain variance refers to shifts in feature distributions between training and testing datasets. For modern IoT systems, the domain variance is often caused by multiple interacting factors. Traditional model transfer methods, which address single-factor shifts, struggle in these complex scenarios. To overcome this limitation, we present VMA, a model transfer framework that decomposes multi-factor distribution shifts into manageable single-factor problems, using a multi-task transfer learning approach. VMA leverages the physical principles of sensors to enhance robustness and efficiency in real-world IoT sensing applications.The proposed methodologies are evaluated with real-world IoT sensing data, demonstrating significant improvements in the robustness to the data shift and accuracy of human information inference.