Mahmoud Youssef
AI Researcher
PhD Researcher · Wrocław University of Science and Technology
I work on generating synthetic data that is safe to share and still trustworthy to learn from. My research brings together differential privacy, generative modelling, and continual learning, so that models trained on synthetic data keep making the right decisions as the underlying data changes.
Current direction Privacy-Preserving Synthetic Data Generation for Trustworthy and Adaptive Machine Learning, PhD in Information and Communication Technology, Wrocław University of Science and Technology.
Wrocław, Poland
Three questions, one problem
Most of my work returns to a single difficulty: we increasingly need to learn from data we are not allowed to look at. Synthetic data is the usual answer, but generating it privately is only half the problem. The other half is whether anything learned from it still holds.
-
01
Privacy-Preserving Generative Modelling
Can we generate data that is provably private and still statistically useful enough to train on?
Explore the theme -
02
Decision-Faithful and Trustworthy Learning
Does a model trained on synthetic data still make the same decisions — and can it explain them?
Explore the theme -
03
Adaptive Learning Under Distribution Shift
Real data streams move. How do generators and the models trained on them stay correct when they do?
Explore the theme
What I am working on now
My doctoral research continues the line of work begun in my master's thesis, but widens it in two directions: beyond a single tabular dataset toward multi-modal data streams, and beyond a fixed distribution toward data that keeps moving. These are the questions I am currently trying to answer.
- Can a differentially private generator stay faithful to the decision boundary, rather than only to the marginal distributions it is easiest to measure?
- What happens to those guarantees when the data stream is non-stationary and the generator has to keep learning?
- How should synthetic data be evaluated, so that the evaluation itself is trustworthy?
Systems and studies
Work that produced a system or an empirical result rather than a paper — and, in one case, both.
At a glance
Get in touch
I am glad to hear from researchers working on privacy, synthetic data, or trustworthy machine learning — whether that is a question about a paper, a possible collaboration, or an idea worth arguing about.
MahmoudGamalfl8@outlook.com