Background

I am interested in the domain of machine learning and the ways it works (and by extension, the ways it doesn’t work). I find mathematics (especially discrete mathematics and linear algebra) and theory very cool, and I have always tried my best to formalise my intuition using these as a backbone.

Experience

As of yet, my research experience includes a lot of reading, understanding and some minor projects such as a study investigating improvements on attacks against LLM watermarks. I’m currently working on one major project related to backdoor vulnerabilities in model merging setups. A preprint for the same is currently in preparation.

Research Interests

I am actively looking to explore problems at the intersection of theoretical ML and security. These include:

• Security & Privacy in LLMs
  LLMs are absolutely ubiquitous today, and unfortunately (or fortunately, depending on who you are) they are not foolproof. My interest is in investigating why and how they can fail, and thence coming up with new and innovative ways to induce more such failures. Currently I’m trying to get into the theoretical side of this domain as well, to understand if we can give any concrete-ish guarantees on the robustness of these models.

• Adversarial Machine Learning
  I always love adversarial examples because of their elegant and simplistic goal: to break things. I am interested in (1) improving said adversarial attacks to break things better and (2) trying to establish defenses against such attacks (ideally using mathematically grounded methods).

• Interpretability/Explainability
  Machine learning models (especially the big ones) are frustratingly opaque. Revealing their inner secrets in the form of both a priori reasoning and post hoc analysis is something that is of great interest to me.

Note: I am also open to adjacent problems across the broader machine learning and information security umbrellas!

I am always open to collaboration and discussing/exploring new ideas. Feel free to drop me an e-mail!