UIS Lab — Welcome

What the Lab Does

The UIS Lab pioneers intelligent and human-centered sensing technologies to transform health and well-being. We design trustworthy, privacy-aware, and robust AI systems that can sense, interpret, and respond to the complexities of human life—advancing early detection, personalized interventions, and scalable health solutions. By bridging engineering, computing, and biomedical science, we address challenges of distribution shifts, algorithmic fairness, AI ethics, and reliability with optimization methods, explainable AI solutions, and mathematical guarantees for trustworthy performance. Our research advances LLM and foundation model solutions tailored for human sensing—overcoming limited training data, capturing human heterogeneity, and building reasoning that aligns with health experts to deliver reliable, meaningful, and seamlessly integrated healthcare decisions.

Our Projects

Human Heterogeneity Invariant Stress Sensing

Human heterogeneity remains a persistent challenge in wearable-based stress sensing. In this paper, we introduce a pruning-based framework designed to improve AI model generalizability, achieving a 7.8-percentage-point increase in stress detection accuracy under fully unseen subjects and environments.

Collaborate

Reading between the heat.(Co-teaching)

Thermal stress sensing often struggles at varying distances and angles, degrading accuracy. We developed ThermaStrain, a co-teaching framework that uses wearable EDA signals during training to teach the thermal modality richer stress features. This enables robust, contactless stress detection at test time—improving accuracy by over 9% across diverse distances and perspectives.

Collaborate

Multimodal LLMs and Trustworthy AI

We build evaluation frameworks for multimodal LLMs, and design multimodal systems that integrate sensor data for health monitoring.

Evaluation frameworks for LLM/VLM for multimodal inputs
Designing multimodal foundation models
Trustworthy AI with theoretical guarantees

Collaborate

Responsible Generative AI for Health Applications

We leverage diffusion and related generative models to create high-utility synthetic health data with privacy-preserving mechanisms, balancing data utility with strict privacy constraints so researchers can work with realistic datasets without exposing identifiable human data.

Collaborate

Everyday Device Sensing for Psychological & Physiological States

We use signals from everyday devices (phones, wearables) for unobtrusive sensing of stress, social anxiety, and affect, emphasizing robustness and privacy-first deployment to keep systems effective and trustworthy at scale.

Collaborate

Cyber-physical Systems for Cattle wellbeing on Diary Farms

Our project develops cyber-physical systems that use sensor networks to monitor the health, behavior, and social interactions of dairy cows in real farm environments. By combining localization, motion, and thermal sensing, we aim to understand stress, wellness, and productivity at both the individual and herd level. These tools can help farmers manage animal health more effectively while also advancing broader research in sensing and animal science. Alongside system development, we are addressing key challenges such as energy efficiency, network reliability, and long-term deployment in harsh farm conditions.

Collaborate

Curvature-Aligned Federated Learning (CAFe): Fairness without Demographics

We developed CAFe, a federated learning approach that achieves fairness without relying on sensitive demographic attributes. By aligning loss-landscape curvature during local training and across clients, CAFe balances fairness and performance in decentralized human-sensing applications where bias factors are unknown or latent. Validated through theory, experiments, and real-world deployment on resource-constrained devices, CAFe demonstrates both practical feasibility and strong potential for privacy-preserving, equitable AI at the edge.

Collaborate

Ethical AI for Healthcare

AI is rapidly transforming healthcare, but many systems focus on overall accuracy while overlooking ethical considerations. A model may perform well on average yet still deliver unequal or harmful outcomes for certain patients. This project explores how to evaluate and guide AI systems to be not only fair but also ethical—promoting well-being, avoiding harm, ensuring equity, and fostering trust—without relying on sensitive demographic attributes.

Collaborate

Meet the People

Supervisor

Prof. Asif Salekin (PI)
LinkedIn Scholar

Leads UIS Lab. Research interests: human‑centered AI, mobile/wearable sensing, fairness, robustness, interpretability, privacy & reliability.

Ph.D. Students

Yi Xiao LinkedIn

Stress sensing and multimodal learning; co‑teaching and thermal signals for non‑intrusive inference.

Harshit Sharma

Responsible and trustworthy AI in ubiquitous sensing applications, with an emphasis on model interpretability, fairness, and privacy.

Shaily Roy

Fairness‑aware federated learning; cluster‑aware personalization for digital health.

Sachin Deb

Ubiquitous computing and machine learning for health applications.

Graduated Ph.D. Students

Jingyu Xin — Secure & trustworthy human‑centric sensing; defenses against data‑injection attacks.
Brian Philip Testa — Emotional privacy and adversarial audio for speech emotion systems.
Sawinder Kaur — Verified robust compressed networks; context‑robust personalization.

M.S. & Undergraduate Researchers

Laxman Kumar, Guillermo Ramon Sanchez, Lohith Srikanth Pentapalli, Anusha Prakash, Jeffrey Ching, Mudit Dham, Fang Ziqi, Avery T. Gump, Peter Cheng — Contributing to sensing pipelines, modeling, and evaluation.