Alireza Dehghanpour Farashah

PhD Student in Computer Science

McGill University and Mila, Quebec AI Institute

Montréal, Canada

I am Alireza, a PhD student in Computer Science at McGill University and Mila, Quebec AI Institute.

My research lies at the intersection of machine unlearning and concept erasure, with a particular focus on large-scale vision-language and text-to-image models. I study how undesired concepts, such as sensitive, biased, or copyrighted information, can be selectively removed from trained models while preserving their overall utility.

My academic background includes a Bachelor’s degree in Computer Engineering from Sharif University of Technology, where I developed a strong foundation in machine learning and computer vision. Building on this background, my current work investigates robustness, reliability, fairness, and safety in modern generative models, with the goal of developing principled methods for controllable and responsible AI systems.

selected publications

Blacksmith: Fast Adversarial Training of Vision Transformers via a Mixture of Single-step and Multi-step Methods (arXiv)

Mahdi Salmani*, Alireza Dehghanpour Farashah*, Mohammad Azizmalayeri, and 4 more authors

2023

Bib

@misc{salmani2023blacksmith,
  title = {Blacksmith: Fast Adversarial Training of Vision Transformers via a Mixture of Single-step and Multi-step Methods <a href="https://arxiv.org/abs/2310.18975">(arXiv)</a>},
  author = {Salmani*, Mahdi and Farashah*, Alireza Dehghanpour and Azizmalayeri, Mohammad and Amiri, Mahdi and Eslami, Navid and Manzuri, Mohammad Taghi and Rohban, Mohammad Hossein},
  year = {2023},
  eprint = {2310.18975},
  archiveprefix = {arXiv},
  primaryclass = {cs.CV},
}