Mahdi Haghifam

Email(preferred): haghifam.mahdi@gmail.com Email: mhaghifam@ttic.edu

About Me

I am an ML researcher interested in developing principled algorithms for reliable and efficient AI/ML. My work spans foundations of ML — including optimization and generalization — as well as privacy-preserving ML, privacy auditing, and memorization in learning systems. Recently, I've been developing principled inference-time algorithms to address challenges in LLM safety and efficiency.

I'm currently a Research Assistant Professor at TTIC. Previously, I was a Distinguished Postdoctoral Researcher at Northeastern University, working with Jonathan Ullman and Adam Smith. I completed my Ph.D. in Machine Learning (thesis) at the University of Toronto / Vector Institute, advised by Daniel M. Roy. I also hold B.Sc. and M.Sc. degrees in Electrical Engineering from Sharif University of Technology.

I've worked in industry at Google DeepMind (with Thomas Steinke), where I built second-order differentially private optimization methods achieving 10–40× speedups over standard baselines, and at ServiceNow Research (with Gintare Karolina Dziugaite), where I developed data-dependent generalization estimation tools for deep networks. See details here.

Recognitions include a Best Paper Award at ICML 2024 (top 10 of 10,000 submissions), a Simons Institute–UC Berkeley Research Fellowship, and Top Reviewer at NeurIPS 2021 and 2023.

Research Overview and Selected Papers

I work on foundations and algorithms for machine learning, especially in settings where rigorous guarantees are important. My research aims to develop principled methods for modern ML systems that are both theoretically grounded and practically useful. Recently, my work has focused on reliable and efficient AI, spanning privacy-preserving learning, memorization, model stealing in large language models, and inference-time methods for controlling and verifying model behavior.

• The Distillation Game in LLMs: Adaptive Attacks & Efficient Defenses
  M. Haghifam*, Y. Allouah*, R. Shokri, S. Koyejo
  Manuscript available upon request, 2026

• Black-Box Data Reconstruction via List Decoding: The Necessity of List Memorization for Learning
  G. Brown, M. Haghifam, E. Linder, R. Livni, G. Thampakkul, N. Weinberger
  Under prepreation, 2026

• The Sample Complexity of Membership Inference and Privacy Auditing
  M. Haghifam, A. Smith, J. Ullman
  Pre-print, 2025

• Information Complexity of Stochastic Convex Optimization: Applications to Generalization and Memorization |  [Talk video at ICML'24]
  I. Attias, G. K. Dziugaite, M. Haghifam, R. Livni, D. M. Roy (alphabetic ordering)
  International Conference on Machine Learning (ICML), 2024 (Oral, Best Paper Award: Top 10 of 10,000 submissions)

• Private Geometric Median |  [Talk video] |  [Code]
  M. Haghifam, T. Steinke, J. Ullman
  Advances in Neural Information Processing Systems (NeurIPS), 2024

• Faster Differentially Private Convex Optimization via Second-Order Methods |  [Code]
  A. Ganesh, M. Haghifam, T. Steinke, A. Thakurta (alphabetic ordering)
  Advances in Neural Information Processing Systems (NeurIPS), 2023

• Sharpened Generalization Bounds based on Conditional Mutual Information and an Application to Noisy, Iterative Algorithms |  [Talk video] |  [Code]
  M. Haghifam, J. Negrea, A. Khisti, D. M. Roy, G. K. Dziugaite
  Advances in Neural Information Processing Systems (NeurIPS), 2020

Industry Internship Experience

Contact Me!

Feel free to reach out if you'd like to discuss research ideas. Also, I'm happy to offer guidance and support to those applying to graduate programs, especially individuals who might not typically have access to such assistance