IARPA HIATUS: Human Interpretable Attribution of Text using Underlying Structure
Sep 2025 – Present
USC ISI · UMD · UMich · UBirmingham
- Formalized evaluation protocols for user and attribute privacy leakage in speech.
- Developing a dataset and benchmark for authorship attribution from spoken language.
- Designing methods for speaker anonymization across both acoustic and linguistic modalities.
Speaker Verification
Speaker Anonymization
DP
Visual Privacy Understanding
Sep 2024 – Present
USC · Amazon
- Received a gift award from the USC-Amazon Center on Secure and Trusted Machine Learning.
- Developed a taxonomy of visual privacy risks grounded in legal frameworks.
- Evaluated VLMs on privacy recognition and compositional reasoning.
- Designing deployable VLM-as-a-judge models for privacy severity assessment, along with a dataset of compositional privacy risks.
- Building a privacy advisory agent to detect, localize, and sanitize sensitive visual content for regulatory compliance.
Vision Language Models
Benchmarking
Reasoning
Agentic AI
Horizon Europe - PILLAR-Robots: Purposeful Intrinsically motivated Lifelong Learning Autonomous Robots
Jan 2024 – Jul 2024
Athena RC · UDCoruna · CNR · Sorbonne · AI2Life · PAL Robotics
- Developed robotic simulation pipelines
- Enabled the robot to learn useful environmental information under minimal interaction during the exploration phase.
- Investigated domain adaptation techniques to transfer learned behaviors from simulation to real-world robotic settings.
Vision Language Models
Robotics
Lifelong Learning
Enhancing Vision-Language Pre-training
Nov 2022 – Jul 2022
NTUA · UT Austin
- Proposed a multimodal extension of contrastive vision–language models by incorporating dialogue as an additional modality.
- Generated synthetic dialogue data to augment image–text datasets with richer contextual descriptions.
- Developed a domain adaptation method to adapt models to dialogue inputs via a dual contrastive objective while preserving generalization.
Vision Language Models
Self-supervised Learning
Domain Adaptation
Synthetic Data