Research

Privacy-Preserving Data Publishing with Agentic AI

As data-driven decision-making becomes central to fields like healthcare, finance, and public policy, protecting individual privacy while keeping data useful is one of the most pressing challenges in modern computing.

My research focuses on Privacy-Preserving Data Publishing (PPDP) — the science of releasing data in ways that prevent individuals from being identified, without destroying the data’s value. I work with established anonymization techniques such as k-anonymity, and am extending this work to richer privacy guarantees including l-diversity and t-closeness.

The novel direction of this project is the development of an agentic AI system that automates the anonymization pipeline while keeping a human expert at the centre of the decision. A user uploads a dataset, selects the attributes to protect, and the system produces a structured report comparing multiple anonymized versions — across algorithms, utility metrics, and privacy trade-offs. A domain expert then reviews the report and selects the best outcome based on their contextual knowledge. This human-in-the-loop design reflects a core belief: that algorithmic outputs should inform expert judgment, not replace it.

The long-term goal of this work is to propose v-difference, a new privacy model designed to improve on the limitations of existing approaches, formally defined and empirically validated within the AI system.

Bias and Fairness in Computer Vision with Agentic AI

As computer vision systems are deployed in increasingly high-stakes environments — from law enforcement to autonomous vehicles — understanding where and why these models fail is as important as making them accurate.

My research in this area investigates multiple dimensions of bias in vision models. The first is acquisition bias: how the source and quality of an image systematically affects model performance, independent of the subject being photographed. Drawing on my graduate background in signal and image processing, I study how cameras with fundamentally different characteristics — such as body-worn cameras versus high-resolution fixed cameras — produce performance gaps that standard benchmarks rarely capture. The second dimension extends this to gender and racial bias, using data collected across acquisition conditions and comparing model behaviour against existing benchmark datasets to understand how demographic disparities interact with image quality.

The work begins with traffic sign detection, a domain with well-established datasets and clear ground truth, before extending to broader image recognition tasks. As with my privacy research, the goal is not only to measure these gaps but to build an agentic AI system that automates bias auditing across models and data sources, presenting findings in a structured report for a human reviewer to interpret and act on.