Mahdi Haghifam

Email(preferred): [email protected] Email: [email protected]

About Me

I am a ‪research assistant professor at the Toyota Technological Institute at Chicago (TTIC)‬. I was previously a Distinguished Postdoctoral Researcher at Khoury College of Computer Sciences at Northeastern University, fortunate to be working Jonathan Ullman‬ and Adam Smith‬. I completed my Ph.D. in Machine Learning (thesis‬) at University of Toronto/ ‪Vector Institute‬‬ where I was fortunate to be advised by ‪Daniel M. Roy‬. I also received my B.Sc. and M.Sc. degrees in Electrical Engineering from Sharif University of Technology.

During my Ph.D., I had a great time working in industry as an intern at Google DeepMind (mentored by Thomas Steinke) where I worked on the algorithm design for privacy-preserving optimization and ServiceNow Research (mentored by Gintare Karolina Dziugaite) where I worked on generalization in deep learning. See details here.

Recognitions of my work include a Best Paper Award at ICML 2024, Simons Institute-UC Berkeley Research Fellowship, as well as several honors for graduate research excellence from University of Toronto, including the Henderson and Bassett Research Fellowship and the Viola Carless Smith Research Fellowship. Additionally, I was recognized as a top reviewer at NeurIPS in 2021 and 2023.

Research Overview and Selected Papers

My research focuses on the foundations of trustworthy machine learning and principled algorithm design. The central goal of my work is to address practical challenges in ML by developing tools and algorithms with rigorous theoretical guarantees that assess and ensure validity. This is crucial for building trustworthy systems in high-stakes applications, where balancing responsible deployment with strong empirical performance is essential. Currently, I am extending my work to address challenges in AI safety, developing principled post-training mechanisms for reliability, efficiency, and alignment.

Generalization in Machine Learning:

• Information-Theoretic Generalization Bounds for SGLD via Data-Dependent Estimates |  [Code]
  J. Negrea*, M. Haghifam* , G. K. Dziugaite, A. Khisti, D. M. Roy
  Advances in Neural Information Processing Systems (NeurIPS), 2019

• 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

• Towards a Unified Information-Theoretic Framework for Generalization  |  [Talk video]  
  M. Haghifam , G. K. Dziugaite, S. Moran, D. M. Roy.
  Advances in Neural Information Processing Systems (NeurIPS), 2021 (Spotlight, <3% of submissions)

Memorization and Privacy Attacks:

• 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)

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

• On Traceability in Stochastic Convex Optimization
  S. Voitovych*, M. Haghifam* , I. Attias, G. K. Dziugaite, R. Livni, D. M. Roy
  Advances in Neural Information Processing Systems (NeurIPS), 2025 (Spotlight, <3% of submissions)

Differential Privacy:

• 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

Industry Internship Experience

Google DeepMind| Research Intern
Mountain View| September 2022 – December 2022
Mentors: Thomas Steinke
- Developed second-order differentially private optimization method achieving 10–40× wall-clock speedups over DP-SGD baselines at comparable accuracy/privacy
- Resulted in publications at NeurIPS 2023 (link) (code) and ICML 2023 (link)

ServiceNow Research | Research Intern
Toronto | November 2020 – March 2021
Mentor: Gintare Karolina Dziugaite
- Studied the connections between different generalization approaches in ML
- Resulted in publication in NeurIPS 2021 (Spotlight) (link)

ServiceNow Research | Research Intern
Toronto | February 2019 – May 2019
Mentor: Gintare Karolina Dziugaite
- Proposed data-dependent generalization estimates for noisy SGD / SGLD using gradient disagreement. NeurIPS 2019 (link)
- The proposed method achieves for the first time non-vacuous generalization bound in various modern ML setups.
- Built generalization-prediction tools in TensorFlow for CNNs and MLP on image classification tasks. (code)

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