Research
My research focuses on trustworthy multimodal AI, particularly large vision-language models, with an emphasis on privacy-aware reasoning, adversarial robustness, and long-horizon visual understanding.
I study how vision-language models can remain useful while reasoning about privacy-sensitive visual information in a more controlled and reliable way.
I study both adversarial attacks and defenses for vision and vision-language models, with a focus on understanding model vulnerabilities and improving robustness.
I study how multimodal systems can understand and reason over extended egocentric video histories while maintaining useful and privacy-aware behavior.
Research Questions
Each question maps to one of the three research areas.
How can LVLMs reason about scenes without leaking sensitive identity, location, relationship, or behavioral information?
What attacks transfer across encoders and models, and what defenses preserve accuracy without destroying useful visual content?
How can multimodal systems handle long-horizon and egocentric VideoQA while balancing utility, privacy, and reliability?
Research Areas
These are public-facing project areas. Specific unpublished titles, internal figures, experimental results, and submission details are intentionally left out until they are public.
Privacy boundaries, information disclosure, privacy leakage in multimodal reasoning, and privacy-utility trade-offs.
Transferable attacks across models and vision encoders, adversarial purification, detection, and robust generalization.
Long-horizon and egocentric VideoQA, retrieval and reasoning over extended visual context, and utility/privacy trade-offs.
Methods & Tools
The technical stack combines research engineering, ML experimentation, and software development experience.