I work on Responsible and Trustworthy AI — developing methods and frameworks to ensure AI systems are private, fair, robust, and transparent. My work spans the design, auditing, and application of trustworthy AI systems, with a focus on high-impact domains.
Sensitive datasets in domains such as genomics, healthcare, and finance contain enormous research value but often cannot be openly shared due to privacy risks. Synthetic data offers a promising path forward by generating artificial datasets that preserve the characteristics of real data while reducing the risk of exposing individuals. I view synthetic data as an important infrastructure for enabling open science and collaborative research in settings where access to real data is limited.
My research has focused on developing privacy-preserving synthetic data generation methods, particularly when data is distributed across institutions that cannot share their sensitive real data. I have explored approaches that combine differential privacy, secure multiparty computation, and generative models to enable collaborative synthetic data generation while protecting sensitive information. My work has also examined the privacy risks of synthetic data through auditing and attack-based evaluation, highlighting the importance of understanding both the capabilities and limitations of synthetic data.
Moving forward, I aim to extend this direction toward more complex and realistic synthetic data generation settings. Important open questions include how to generate multimodal synthetic datasets where text, images, and tabular information may reveal different types of sensitive information; how to generate longitudinal data that preserves meaningful relationships over time; and how to understand the tradeoffs among privacy, utility, fairness, and robustness when generating synthetic datasets for real-world applications.
AI systems are increasingly deployed in high-impact settings where failures in privacy, fairness, robustness, or transparency can have significant consequences. Building trustworthy AI requires not only designing systems with desirable properties, but also developing methods to evaluate, verify, and assure that these properties are maintained throughout the AI lifecycle.
Building on my experience in privacy-preserving AI, I investigate methods for auditing and assuring trustworthy AI systems. My prior work has examined privacy risks in synthetic data through privacy attacks and red-teaming approaches, highlighting the importance of evaluating whether privacy guarantees hold in practice. I have also explored constructive approaches for trustworthy AI, including privacy-preserving fairness auditing and algorithmic recourse mechanisms that enable evaluation and explanation while protecting sensitive information.
Moving forward, I aim to broaden this direction toward comprehensive AI assurance frameworks that address multiple dimensions of trustworthiness. With the emergence of foundation models and agentic AI systems, I am interested in developing methods to develop new methods of auditing systems.
As AI systems become increasingly integrated into human decision-making, ensuring that their behavior is appropriate requires more than optimizing technical performance. Many existing approaches provide formal guarantees for individual aspects of trustworthy AI, but they often do not fully capture the social, situational, and domain-specific context in which AI systems operate. For example, whether an AI decision is fair, whether an explanation is meaningful, or whether information use is appropriate can depend strongly on the context in which the system is deployed.
Building on my background in privacy-preserving AI, I am interested in exploring how contextual information can help bridge formal guarantees with human expectations. My previous work on differential privacy, secure multiparty computation, fairness, and trustworthy AI provides a foundation for investigating how notions of privacy, fairness, transparency, and faithfulness can be grounded in the contexts where AI systems are used.
This emerging research direction explores fundamental questions about context-aware trustworthy AI: How should notions of privacy, fairness, and faithfulness, evolve for foundation models and agentic AI systems that operate across multiple contexts?
AI systems have the potential to transform critical domains such as healthcare and education, but their adoption requires addressing challenges related to privacy, fairness, transparency, and trust. I have contributed in applying trustworthy AI methods to real-world problems through interdisciplinary collaborations such as conversational AI systems that support patients through challenging treatment journeys, synthetic data for rare diseases, and AI-assisted learning systems.
Moving forward, I aim to continue developing trustworthy AI systems together with domain experts, where real-world challenges inform the methods we create.