projects

Research Goal

My research sits at the intersection of representation learning and model reliability in deep learning, with a focus on how complex neural systems encode, reason over, and sometimes fail to disentangle fine-grained visual information in security-critical settings. I pursue two complementary threads:

• Generative face demorphing & biometric security. Morph attacks let a single forged face match multiple identities and threaten face recognition systems. I build diffusion- and GAN-based reconstruction models, benchmarks, and evaluation protocols that not only detect morphs but recover the constituent identities — providing forensic evidence and strengthening downstream morph attack detection.
• Multimodal large language models (MLLMs). I study how MLLMs encode and reason over visual information, addressing two questions: (a) how to optimize MLLMs for multi-image reasoning without extensive human annotation, and (b) how to leverage their internal representations to learn richer features for downstream retrieval and generation tasks.

The unifying goal is trustworthy visual intelligence: models that are accurate, interpretable, and dependable when the stakes are high.