Machine Unlearning
Methods for selectively removing unwanted information from language models while preserving coherence and general utility.
Research
Making language models safer to trust.
My work examines unwanted memorization, privacy leakage, targeted forgetting, and adversarial behavior in large language models.
Focus areas
Three connected problems
I approach trustworthy AI across the model lifecycle: understanding what models retain, measuring how that information can leak, and developing safer mitigation strategies.
Privacy & Memorization
Membership inference attacks and model behavior analysis for understanding when training data can be detected or exposed.
Safety & Alignment
Adversarial in-context learning, jailbreaking, poisoning, and safety alignment in large language and reasoning models.
Research questions
From model behavior to dependable safeguards.
The goal is to develop AI systems whose privacy and safety properties are measurable, controllable, and robust.
- How can a model forget targeted knowledge without losing broad utility?
- What model signals reveal memorization or training-set membership?
- How do adversarial contexts and post-training affect safety behavior?
- How can mitigation strategies remain robust across model behaviors?
Experience
Research & teaching
-
Aug 2025–Present
Graduate Research Assistant
Wayne State University
Research on LLM safety, privacy risks, unlearning, membership inference, and model behavior analysis.
-
Aug 2024–Aug 2025
Graduate Teaching Assistant
Wayne State University
Introduction to Machine Learning, Graduate Seminar, and Operating Systems.
-
Aug 2023–Aug 2024
Graduate Research Assistant
Wayne State University
Research related to large language models and AI safety.
-
2021–2022
Teaching Assistant
University of Tehran
Graph Theory.
Methods & tools
Technical background
Machine learning
Frameworks & libraries
Programming