STAI CDT PhD student Atri Sharma reached an important milestone this summer, presenting a poster at his first conference, the 41st Conference on Uncertainty in Artificial Intelligence (UAI 2025) in Rio de Janeiro, Brazil. The paper, ‘Learning Robust XGBoost Ensembles for Regression Tasks,’ explores new approaches to making AI systems more resilient against adversarial attacks.
We spoke to Atri about the paper and his highlights from UAI 2025.
Congratulations on this achievement! Could you explain what your paper is about and how it contributes to safe and trusted AI?
This work introduces a new way to train decision-tree models so they are more robust against “adversarial attacks.” These attacks involve tiny, carefully chosen changes to a model’s input that can trick it into making the wrong decision. Decision-tree ensembles are widely used in sensitive, real-world settings — for example, to detect fraudulent credit-card transactions. They are usually accurate, but even a very small tweak to the input (like lowering a transaction amount by just £0.01) can cause the model to miss fraud entirely. Worryingly, attackers don’t need much information about how a model works to find these weaknesses.
We propose a method for training decision-tree ensembles that makes them much more resistant to these kinds of manipulations. Unlike previous approaches, which were slow and often tailored to specific tasks, our method is fast, broadly applicable, and provides mathematical guarantees of robustness. This can help make AI systems safer, more trustworthy, and harder for malicious actors to exploit.
That sounds fascinating. What were some of the highlights from your conference experience?
The conference had several memorable highlights, both academically and personally. To begin with, it was held in Rio de Janeiro, and getting to experience its vibrant culture, stunning beaches, and overall energy made the trip genuinely unforgettable. The setting added a sense of adventure to the whole experience.
Academically, I was exposed to several areas that I hadn’t engaged with much before, particularly topics in causality. Learning about these new perspectives broadened my understanding of the field and gave me fresh ideas about how different areas of machine learning connect and complement each other.
The conference itself was relatively small, which created a very intimate and welcoming atmosphere. Because of this, I was able to have in-depth conversations with many of the attendees and organisers, rather than just brief exchanges. One of the most enjoyable moments was a samba night organised for the entire group; sharing an evening of music and local culture with everyone was a real highlight.
Conferences are such an important part of the PhD experience, offering opportunities for networking and exposure to new ideas. How has this conference influenced your research and your broader PhD journey?
My PhD is going well overall, but like many long research projects, it can sometimes feel all-consuming. That’s why attending this conference, which was my first – felt so meaningful. I was exposed me to a much wider range of ideas and research directions than I normally, and hearing about projects that were far removed from my own helped me broaden my perspective and appreciate the diversity in the field.
What I found particularly valuable was seeing the field through so many different lenses – from students exploring new research directions, to industry researchers tackling real-world challenges, to academics thinking deeply about long-term scientific questions. These conversations helped me understand not only what people are working on, but also why those problems matter to them. As someone still early in my research career, that variety of viewpoints was eye-opening and motivating.
Overall, this first conference experience showed me how important it is to stay connected to the wider community. It broadened my thinking about where my own thesis could go and made the PhD process feel less isolated and much more inspiring.
Congratulations again! It sounds like a memorable first conference experience, and one that has set an exciting tone for the rest of your doctoral journey. We look forward to hearing more about your research and your contributions to safe and trustworthy AI.
You can read the full paper here: Sharma, A. V., Kouvaros P. & Lomuscio, A. (2025). Learning Robust XGBoost Ensembles for Regression Tasks. In Proceedings of the 41st Conference on Uncertainty in Artificial Intelligence (UAI 2025).