We introduce 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 weight updates and re-weighting high-/low-energy directions, MonoSoup reduces ensembling cost from O(K) to O(1) while preserving the generalization benefits of weight-space averaging.
@inproceedings{abdollahpoorrostam2026monosoup,title={Model Soups Need Only One Ingredient},author={Abdollahpoorrostam, Alireza and Dimitriadis, Nikolaos and Hazimeh, Hussein and Frossard, Pascal},booktitle={International Conference on Machine Learning (ICML)},year={2026},}
ICLR
A General Framework for Black-Box Attacks under Cost Asymmetry
Mohammadreza Salmani, Alireza Abdollahpoorrostam, and Seyed-Mohsen Moosavi-Dezfooli
In International Conference on Learning Representations (ICLR), 2026
A general framework for black-box adversarial attacks under asymmetric query costs, introducing Asymmetric Search and Asymmetric Gradient Estimation (AGREST) algorithms.
@inproceedings{salmani2026agrest,title={A General Framework for Black-Box Attacks under Cost Asymmetry},author={Salmani, Mohammadreza and Abdollahpoorrostam, Alireza and Moosavi-Dezfooli, Seyed-Mohsen},booktitle={International Conference on Learning Representations (ICLR)},year={2026},}
Preprint
Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning
Alireza Abdollahpoorrostam, Ehsan Sharifian, Buse Şen, and 2 more authors
We recast adversarial training through the lens of optimal transport. Using Brenier maps, we derive an OT-geometric formulation that yields principled, provably robust learning algorithms.
@unpublished{abdollahpoorrostam2026brenier,title={Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning},author={Abdollahpoorrostam, Alireza and Sharifian, Ehsan and Şen, Buse and Cuturi, Marco and Kuhn, Daniel},note={Under review},year={2026},}
2024
NeurIPS
SuperDeepFool: A New Fast and Accurate Minimal Adversarial Attack
Alireza Abdollahpoorrostam, Mahed Abroshan, and Seyed-Mohsen Moosavi-Dezfooli
In Advances in Neural Information Processing Systems (NeurIPS), 2024
SuperDeepFool is a fast, accurate minimal-l_2 adversarial attack derived from a geometric reformulation of the DeepFool iteration. It computes provably small perturbations with state-of-the-art quality and runtime.
@inproceedings{abdollahpoorrostam2024superdeepfool,title={SuperDeepFool: A New Fast and Accurate Minimal Adversarial Attack},author={Abdollahpoorrostam, Alireza and Abroshan, Mahed and Moosavi-Dezfooli, Seyed-Mohsen},booktitle={Advances in Neural Information Processing Systems (NeurIPS)},year={2024},}
Workshop
In Search of the Successful Interpolation: On the Role of Sharpness in CLIP Generalization
@inproceedings{abdollahpoorrostam2024sharpness,title={In Search of the Successful Interpolation: On the Role of Sharpness in CLIP Generalization},author={Abdollahpoorrostam, Alireza},booktitle={NeurIPS 2024 AdvML-Frontiers Workshop},year={2024}}
Workshop
Unveiling CLIP Dynamics: Linear Mode Connectivity and Generalization
Alireza Abdollahpoorrostam, Amartya Sanyal, and Seyed-Mohsen Moosavi-Dezfooli
In ICML 2024 Foundation Models in the Wild Workshop, 2024
@inproceedings{abdollahpoorrostam2024clipdynamics,title={Unveiling CLIP Dynamics: Linear Mode Connectivity and Generalization},author={Abdollahpoorrostam, Alireza and Sanyal, Amartya and Moosavi-Dezfooli, Seyed-Mohsen},booktitle={ICML 2024 Foundation Models in the Wild Workshop},year={2024}}