Security & robustness
Understand adversarial manipulation, detect backdoors, and build defenses that withstand hostile inputs.
I’m Birhanu Eshete. My research makes AI behavior traceable, its vulnerabilities visible, and its decisions more accountable.

AI Security, Privacy, Safety,
Interpretability, and Ethics
with Human Agency.
NSF CAREER Award
Fulbright U.S. Scholar
USENIX Distinguished Paper

AI’s Promise, Its Perils, and The Future We Must Choose brings this work into a broader public conversation about opportunity, risk, accountability, and the choices that remain in our hands.
Amazon preorder link coming soon.
I work at the intersection of Security, Privacy, Safety, Interpretability, and Ethics with human agency.
I am an Associate Professor of Computer Science at the University of Michigan–Dearborn, where I lead the Data-Driven Security & Privacy Lab and direct the Cybersecurity Center for Education, Research, and Outreach. I am also a faculty affiliate of the Michigan Institute for Data & AI in Society at the University of Michigan, Ann Arbor.
My research identifies, explains, and mitigates security, privacy, safety, transparency, and ethical risks in AI. I develop methods to trace how data and computation shape model behavior, detect adversarial manipulation, and diagnose and repair failures. This work addresses high-stakes applications in autonomous systems, healthcare, finance, and cybersecurity.
My contributions have been recognized with the NSF CAREER Award, the Fulbright U.S. Scholar Award, the Faculty Research Excellence Award from the College of Engineering and Computer Science, and the USENIX Security Distinguished Paper Award. My research appears in leading security, privacy, and AI venues, has been featured in Science, and has contributed to the NIST Trustworthy & Responsible AI Resource Center.
Research and teaching across the United States, Europe, and Africa inform my perspective on who benefits from AI, who bears its risks, and who can shape its development. I earned my PhD at the University of Trento and my M.Sc. and B.Sc. at Addis Ababa University, completed postdoctoral research at the University of Illinois Chicago, and returned to Addis Ababa University as a Fulbright Scholar.
In keynotes, seminars, interviews, and teaching, I connect technical research with the decisions facing researchers, educators, and organizational leaders. I welcome research collaborations on trustworthy AI and speaking invitations on AI security, responsible deployment, education, and the preservation of human agency.
Read my full CVUniversity of Michigan–Dearborn
Addis Ababa University
University of Michigan–Dearborn
University of Illinois Chicago
I develop methods to uncover and reduce AI risks—from poisoned training data to privacy leaks—with provenance as a foundation for understanding how models behave.
Understand adversarial manipulation, detect backdoors, and build defenses that withstand hostile inputs.
Protect sensitive data while making model decisions more interpretable and accountable.
Connect technical assurance with the conditions needed to deploy AI responsibly.
Explore my citation record, h-index, and complete publication history.

Lightweight hardening reduced membership leakage by up to 95%, with at most 3.3% explanation-utility loss on average in the evaluated image benchmarks.

Tracing inference behavior enables targeted repair: modifying just one network layer improved adversarial accuracy by 55% on average in the evaluated case studies.

Following each training sample’s influence on parameter updates exposes clean-label backdoors that can hide behind apparently legitimate labels.
One of the early university courses dedicated to trustworthy AI, taught by Birhanu Eshete at University of Michigan–Dearborn in Winter 2024.
A complete course at the intersection of robustness, transparency, and accountability.
College of Engineering and Computer Science, University of Michigan–Dearborn
U.S. Department of State
U.S. National Science Foundation
USENIX Security Symposium · NAVEX

Front cover
Read the complete preface.
AI’s Promise, Its Perils,
and The Future We Must Choose
A clear-eyed conversation about AI’s possibilities, the harms already here, and the choices still in our hands. Written with a global lens, the book asks who benefits, who bears the risks, and how we preserve human agency.
“The unit of success is not model size.
It is durable human capability.”
Discovery, education, creativity, and development: where AI can expand access to expertise and human capability.
Amazon preorder link coming soon.
BIC Village · DEF CON 30
Promising progress and remaining challenges toward trustworthy machine learning.
Watch talkUniversity of Michigan–Dearborn Reporter
A conversation on the challenges behind reliable and responsible AI.
Read interviewWXYZ Detroit
Practical ways to strengthen cybersecurity beyond the office.
Watch interviewNew Frontiers Podcast
A discussion of machine learning in adversarial environments.
Listen to episodeScience Magazine Podcast
The security challenges facing machine learning systems.
Listen to episodeUSENIX Enigma · San Francisco
How lessons from exploit kits can inform reactive, proactive, and counter-offensive strategies against adaptive cybercriminals.
Watch talk & view slidesShare your event, audience, or collaboration idea.
birhanu@umich.edu