Muhammad Ashiq

About Me

Portrait of Muhammad Ashiq

🔗 LinkedIn
💻 GitHub
📚 Google Scholar

Hello! I am an EECS Ph.D. student at the University of Michigan in the DeepThink Lab, advised by Qing Qu and working closely with Ismail Alkhouri at Los Alamos National Laboratory.

My research interests are in flow-based generative models (e.g. diffusion, flow matching, etc.). In particular, I work on:

  • Trustworthiness: Attacks and defenses on the reliability, privacy, and security of these models.
  • Scientific Applications: Extending these models to solve problems across science and engineering: PDEs, data assimilation, inverse problems, and more.

Previously, I was an undergraduate at the University of Wisconsin-Madison studying mathematics and computer science, working with Grigorios Chrysos on trustworthy ML. I also worked with Yeyu Wang and the Epistemic Analytics lab on learning analytics.

Selected Publications

* denotes equal contribution.

Thumbnail for DDIM vs DDPM paper

Why DDIM Hallucinates More than DDPM: A Theoretical Analysis of Reverse Dynamics

M. H. Ashiq*, S. Arora*, A. N. Harish, I. Kharbanda, H. Y. Tseng, G. G. Chrysos

International Conference on Machine Learning (ICML), 2026.

arXiv PDF

Thumbnail for test-time privacy paper

Inducing Uncertainty on Open-Weight Models for Test-Time Privacy in Image Recognition

M. H. Ashiq, P. Triantafillou, H. Y. Tseng, G. G. Chrysos

Neural Information Processing Systems (NeurIPS) Workshop on Regulatable ML, 2025.

arXiv PDF Code

Thumbnail for arithmetic length generalization paper

Data Augmentations for Arithmetic Length Generalization in Transformers

L. Zhou, M. H. Ashiq, G. G. Chrysos

Neural Information Processing Systems (NeurIPS) Workshop on What Can't Transformers Do, 2025.

Other publications, including all those in learning analytics, can be found at my Google Scholar.

Miscellaneous