Research Area

We develop adversarial in-context learning (ICL) attacks that hijack LLM outputs using imperceptible suffixes and propose defense strategies to restore safety. Our gradient-guided method achieves high stealthiness and transferability, while defenses using clean demos effectively recover model alignment and robustness. 

We design algorithms (ASU) that enable large language models to forget unwanted or sensitive knowledge while preserving overall utility. Our self-distillation framework effectively removes targeted information across various scenarios, achieving strong performance on existing benchmarks. 

We develop generative vision–language frameworks that unify perception and language understanding for grounded spatial reasoning. Our approach combines effective text generation with reasoning-driven segmentation, advancing scene comprehension and accessibility-aware navigation through spatially coherent and interpretable outputs across diverse environments.

We design frameworks that bridge the gap between visual and textual understanding in medical imaging. By aligning structured visual features with clinical language representations, our work enhances cross-modal reasoning and enables more accurate segmentation of anatomical and pathological regions across various modalities. 

We develop accurate, hardware friendly and privacy/power consumption-aware on-device AI systems for natural language processing and medical imaging processing.  

We develop spatial-temporal machine learning models and online algorithms to leverage the geocoded IoT data for learning feature representation, forecast and anomaly detection.  

We design fairness-aware decision support system to address the sampling, social, algorithmic biases in predictive analytics.

We develop medical AI system to automatic interpret medical images with visual aid by exploiting convolutional and recurrent neural networks to extract visual and textual features. 

We use domain knowledge in the format of meta-data to design interpretable feature mapping and learning approaches for explainable and personalized recommendation.  

We apply 4D CNN model and algorithm to learn the key developmental features from fMRI images.

We employ natural language generation models to develop context-aware and location-based conversational agent.