About

Haruka Eshima, a first-year PhD student at Okinawa Institute of Science and Technology co-supervised by Prof. Makoto Yamada and Prof. Amedeo Roberto Esposito.

I am interested in statistical inference in realistic settings (e.g., true distribution is heavy-tailed, access to training data is restricted). I enjoy learning Robust Statistics, Information Theory, Convex Analysis, and Optimal Transport.

My goal is to connect these topics to fundamental limits of security.

• When an ideal source of information is deliberately restricted for privacy or security, what can still be inferred? What utility must be sacrificed to make sensitive inference impossible?

For example,

• Robust mean estimation in a Heavy-tailed setting (current project),

• Information-theoretic limits of Training Data Reconstruction under Differentially Private training,

• Formalizing what LLM security means,

and more …

I would eventually like to connect theoretical analyses with experiments on modern AI systems to identify when and why privacy or security failures arise in practice.

I am drawn to the Socratic idea of the gadfly—using careful questioning to challenge assumptions that have become too comfortable. I hope my research can play a small part in doing this for the systems we increasingly rely on.

I also like to find theories that can predict experimental phenomena in simple neural networks. I analyzed neural network architectures for my Master’s thesis, and connected the theory to experimental results in my first project in PhD.

In my free time, I enjoy Karate club activities, Duolingo (French, Chinese, Latin), and colored-pencil drawing.