Ph.D. Student · Kahlert School of Computing · University of Utah

Soumil Datta

I study how AI systems break, so we can build ones worth trusting.

My research sits at the intersection of machine learning security, generative AI, and robotics: adversarial attacks and defenses for LLMs, VLMs, and diffusion models, and the robustness of learning-based autonomous systems. I'm jointly advised by Daniel Brown (ARIA Lab) and Guanhong Tao (SaLT Lab).

scholar github linkedin orcid cv
person · 0.99 Soumil Datta
// hover to perturb
01

News

  • 2025 Two first-author papers under review: one on sanitizing malicious LLM instructions, and one on evading model IP protections via limited fine-tuning.
  • 2025 "Dataset Poisoning in Behavioral Cloning Policies" accepted at EAI SmartSP 2025.
  • Jan 2025 Presented our generalized attacks on object detectors at HICSS 2025, where it was nominated for Best Paper.
  • Fall 2024 Started my Ph.D. at the Kahlert School of Computing, University of Utah.
  • Jan 2024 🏆 Best Paper Award at HICSS 2024 for EllipScape.
02

Research

/01 · attack & defend

Securing Generative Models

Protecting LLMs, VLMs, and diffusion models from malicious manipulation (adversarial prompts, compromised weights, stolen model IP), and building robust foundations for trustworthy generation.

LLMsVLMsdiffusion
/02 · embodied

Robust Robot Learning

Identifying vulnerabilities in autonomous and robotic systems built on transformer architectures, from dataset poisoning in behavioral cloning to failures in learning-based control.

imitation learningautonomy
/03 · perception

Adversarial Perception

Generalizable, loss-function-based adversarial attacks against state-of-the-art object detection models: understanding what perception systems actually learn, and how it fails.

object detectionattacks
03

Publications

04

Awards & Honors

05

Service

Subreviewer for top venues in security and machine learning:

S&P'25 '26IEEE Symposium on Security and Privacy
NeurIPS'24 '25Neural Information Processing Systems
CCS'25ACM Conf. on Computer and Communications Security
SaTML'26IEEE Conf. on Secure and Trustworthy ML
06

Teaching

Now a full-time research assistant, I previously TA'd Artificial Intelligence and Intro to Machine Learning at the University of Utah, along with a wide range of courses at the University of Mississippi, from Java and data structures to databases and information visualization in R.

Full teaching history →