Raoof
Zare Moayedi

M.Sc. Computer Science Sharif University of Technology

How models learn
and how to improve them

I’m a master’s student in Computer Science at Sharif University of Technology. I previously completed two bachelor’s degrees at Tehran Polytechnic, in Computer Science and Mathematics. I’m curious about how models learn and how understanding them can help us make them work better in practice. I like using both proofs and experiments to work things out. My interests include privacy, robustness, uncertainty, and reasoning, but I often find myself following questions beyond the topics of my current papers.

Google Scholar View my research

Research interests

Choose a topic to explore

A little researcher studies protected data and similarly shaped synthetic points.

Privacy

We study what can be learned under differential privacy and how much data it takes. Our work connects algorithms with lower bounds to understand the trade-off between privacy and accuracy.

Publications & preprints

All 7 publications

Adversarial robustness01

Allies Teach Better Than Enemies: Inverse Adversaries for Robust Knowledge Distillation

Inverse adversarial examples help a smaller student learn a teacher’s robustness.

Scientific ML02

CuPINN: Optimizing PINNs through Curvature Minimization and Residual Landscape Flattening

Penalizing residual curvature to improve how neural networks solve differential equations.

Differential privacy03

Geometry-Adaptive Mechanisms for Private Synthetic Data

Private synthetic data with error guarantees that adapt to intrinsic geometry.

Private sampling04

Minimax-Optimal Private Multi-Sampling from Gaussian Distributions

How much data do we need to privately generate a batch of Gaussian samples?

Covariance estimation05

Fundamental Limits of Private Covariance Estimation under Vertical Partitioning

Private covariance estimation when different parties hold different features.

SGD & robustness06

Understanding When SGD Improves Robustness to Spurious Correlations

Why small batches early in training can reduce reliance on spurious correlations.

Research experience

August 2026 – presentEPFLResearch Intern · E3 Program EPFL logo

I’m working on graph neural networks that use the structure of conventional numerical solvers to predict fluid motion. The goal is to take larger time steps than the solvers can reliably take on their own, while keeping predictions accurate over long periods. Taking fewer steps would make these simulations less expensive.

With Prof. Olga Fink

June 2025 – March 2026University of CambridgeRemote Research Assistant University of Cambridge logo

Developed a geometry-adaptive algorithm for differentially private synthetic data. It uses the data’s intrinsic structure to improve accuracy and reduce computation, with stronger error guarantees than methods governed by the ambient dimension.

Geometry-Adaptive Mechanisms for Private Synthetic Data

With Dr. Amir Reza Asadi

January – June 2025Nanyang Technological UniversityFunded Visiting Research Student

Worked on robust prompt tuning, superalignment, and task arithmetic, studying their connections to adversarial robustness. I examined when adapting a model to a new task or objective preserves its robustness, and when it makes the model easier to attack.

With Prof. Yew-Soon Ong

June 2023 – June 2024Imperial College LondonRemote Research Assistant Imperial College London logo

Worked on transferring adversarial robustness from teacher models to smaller students through robust knowledge distillation. This work was published in IEEE TPAMI.

Allies Teach Better Than Enemies: Inverse Adversaries for Robust Knowledge Distillation

Also worked on two projects in adversarial robustness:

Robust data pruning: studied how data preparation, example selection, and training on pruned data affect adversarial robustness.
Input-space geometry: compared CNNs and Transformers, focusing on how architectural biases affect the connectivity and topology of regions in input space.

With Dr. Seyed-Mohsen Moosavi-Dezfooli