cv
Curriculum vitae — click the icon above to download the latest PDF.
General Information
| Full Name | Alireza Abdollahpoorrostam |
| Position | M.Sc. Researcher, EPFL |
| Affiliations | Chair of Risk Analytics and Optimization (RAO); Signal Processing Laboratory (LTS4) |
| Location | Lausanne, Switzerland |
| alireza.abdollahpoorrostam [at] epfl.ch | |
| Languages | Persian (native), English |
Research Summary
- I am drawn to the mathematical foundations of deep learning — why neural networks fail under adversarial perturbations and distribution shift, and what principled algorithms follow.
- My research spans robustness, generalization, and post-training methods (model merging, model editing) for foundation models, with recent work on the optimal-transport geometry of adversarial training and distributionally robust optimization.
- I aim to derive rigorous guarantees and translate them into deployable algorithms, with work appearing at NeurIPS, ICML, and ICLR.
Education
-
2024 – Present M.Sc. in Communication Systems — Machine Learning track
École Polytechnique Fédérale de Lausanne (EPFL), Switzerland - Chair of Risk Analytics and Optimization (RAO), advised by Prof. Daniel Kuhn — optimal-transport geometry of adversarial training and distributionally robust learning.
- Research Scholar at Signal Processing Laboratory (LTS4), advised by Prof. Pascal Frossard — robust fine-tuning and weight-space merging for vision–language foundation models.
-
2019 – 2024 B.Sc. in Electrical Engineering & Computer Science
Amirkabir University of Technology (Tehran Polytechnic), Iran
Publications
-
Conference Papers
- A. Abdollahpoorrostam, N. Dimitriadis, A. Hazimeh, P. Frossard. Model Soups Need Only One Ingredient. International Conference on Machine Learning (ICML 2026).
- M. Salmani, A. Abdollahpoorrostam, S.-M. Moosavi-Dezfooli. A General Framework for Black-Box Attacks under Cost Asymmetry. International Conference on Learning Representations (ICLR 2026).
- A. Abdollahpoorrostam, M. Abroshan, S.-M. Moosavi-Dezfooli. SuperDeepFool: A New Fast and Accurate Minimal Adversarial Attack. Advances in Neural Information Processing Systems (NeurIPS 2024).
-
Under Review
- A. Abdollahpoorrostam, Ehsan Sharifian, Buse Şen, Marco Cuturi, Daniel Kuhn. Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning. Under review.
-
Workshop Papers & Preprints
- A. Abdollahpoorrostam. In Search of the Successful Interpolation: On the Role of Sharpness in CLIP Generalization. NeurIPS 2024 AdvML-Frontiers Workshop.
- A. Abdollahpoorrostam, A. Sanyal, S.-M. Moosavi-Dezfooli. Unveiling CLIP Dynamics: Linear Mode Connectivity and Generalization. ICML 2024 Foundation Models in the Wild Workshop.
Research Experience
-
Sept. 2025 – Present Graduate Researcher
EPFL — Chair of Risk Analytics and Optimization (RAO) - Advisor: Prof. Daniel Kuhn, Lausanne, Switzerland.
- Bridging optimal transport and adversarial training: OT-geometric formulations of adversarial training via Brenier maps, toward principled, provably robust learning algorithms.
- Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning (with Ehsan Sharifian, Buse Şen, Marco Cuturi, Daniel Kuhn) — under review.
-
Sept. 2024 – Aug. 2025 Graduate Researcher
EPFL — Signal Processing Laboratory (LTS4) - Advisor: Prof. Pascal Frossard, Lausanne, Switzerland.
- Published MonoSoup, a hyperparameter-free, data-free post-hoc method that recovers Model-Soup-level out-of-distribution accuracy from a single fine-tuned checkpoint by applying SVD to per-layer updates and re-weighting high-/low-energy directions; reduces ensembling cost from $\mathcal{O}(K)$ to $\mathcal{O}(1)$. (ICML 2026)
-
2022 – 2025 External Research Collaborator
Trustworthy ML Collaboration with Dr. S.-M. Moosavi-Dezfooli - Remote / Tehran / Lausanne.
- Co-developed SuperDeepFool, a minimal-$\ell_2$ adversarial attack with a geometric reformulation of the DeepFool iteration. (NeurIPS 2024)
- Asymmetric-cost black-box attack, designing Asymmetric Search and Asymmetric Gradient Estimation (AGREST) algorithms for black-box settings. (ICLR 2026)
- Concurrent collaboration with Dr. Amartya Sanyal (U. Copenhagen) on linear mode connectivity of robust fine-tuning. (ICML 2024 FM-Wild Workshop)
Academic Service
- Reviewer — NeurIPS, ICML, ICLR
Technical Skills
| Programming | Python, C/C++, Bash, MATLAB, LaTeX |
| ML Frameworks | PyTorch, JAX |
| Languages | Persian (native), English |
Research Interests
-
Adversarial robustness
- Minimal-norm and geometry-aware adversarial attacks
- Certified defenses and provable robustness
-
Distributionally robust optimization & optimal transport
- Brenier-map formulations of adversarial training
- Wasserstein-DRO for deep learning
-
Foundation models & post-training
- Model merging, model soups, weight-space ensembling
- Robust fine-tuning of vision–language models (CLIP)
- Linear mode connectivity and the loss landscape