Mihnea Ghitu

PhD Student in Trustworthy AI · Department of Computing, Imperial College London

Me by Regent's Canal

Mihnea Ghitu

I am a PhD student in Trustworthy AI in the Department of Computing, Imperial College London, supervised by Dr Matthew Wicker. My research asks how we can obtain provable guarantees about the behaviour of machine learning models, from simple neural networks to vision models and large language models. I work on certification of safety properties, including poisoning robustness, explainability, differential privacy, unlearning and more.

Before my PhD, I completed an MEng in Computing (Machine Learning) at Imperial, graduating with First Class Honours. My master's thesis studied the robustness–privacy trade-off in Bayesian neural networks and was selected as a Distinguished Project.

Alongside research, I have industry experience as an Applied Scientist Intern at Entrust, working on synthetic data generation for face matching, an Electronic Trading Developer Intern at Arctic Lake Technologies, where was involved in creating pricing models and financial data platforms, and a Software Development Engineer Intern at Amazon, where I built large-scale Attribute-Based Access Control Systems.

Feel free to reach out at mihnea.ghitu20@imperial.ac.uk.

News

Sep 2026New preprint: Certification-Based Differentially Private Learning.
Aug 2026Completed my internship at Entrust: outperformed the state of the art in pre-training face-matching models on synthetic data (internal evaluations), and built novel attribute-conditioned generators and training-free benchmarks of data quality.
Apr 2026Started an Applied Scientist internship at Entrust.
Feb 2026New preprint: Towards Poisoning Robustness Certification for Natural Language Generation.
Sep 2025Model Guidance via Robust Feature Attribution published in TMLR.
Oct 2024Started my PhD in Trustworthy AI at Imperial College London with Dr Matthew Wicker.
Jun 2024Graduated with an MEng in Computing (ML) from Imperial with First Class Honours. My Master's Thesis was recognized as a Distinguished Project.