projects
Selected research projects in face demorphing, biometric security, and multimodal large language models.
From research questions to methods
My projects investigate how visual models represent evidence, recover information from ambiguous inputs, and support reliable decisions. Two complementary threads connect this work:
- Generative face demorphing and biometric security. I develop diffusion models, GANs, and identity-preserving decomposition methods to recover contributing identities from a single morphed face. This work also examines generalization to unfamiliar faces and morphing techniques, with evaluation protocols that distinguish identity recovery from input replication.
- Multimodal learning and inference. I study preference optimization for multi-image reasoning, visual token selection under computational constraints, and intermediate representations for detection and attribute-specific retrieval. These projects examine how existing model representations can reduce dependence on additional computation or human annotation.
The projects below present the methods, experiments, and findings behind these directions. For the broader research agenda, visit my research page.
