Published paper at IEEE Big Data 2025

Federated learning lets many parties, called clients, train machine learning models together, and it normally assumes that every client works with every other client. In practice, clients may need to exclude each other for strategic, regulatory or competitive reasons.

In “A Taxonomy and Resolution Strategy for Client-Level Disagreements in Federated Learning”, written with Dr. Ana Oprescu as part of my master’s thesis at the University of Amsterdam, I classify these disagreements and propose a multi-track approach that keeps isolated model update paths, so clients that exclude each other never affect each other’s models. We tested it with an open-source simulation system on MNIST and N-CMAPSS across 34 scenarios. On the server, resolving disagreements took under 1 ms per round in simulation.

I presented it at the 2025 IEEE International Conference on Big Data in Macao. It was my first time outside Europe, so I made the most of it and took my brother with me. I was nervous to present, but in the end it all went fine. Then we spent the rest of the trip in Shenzhen and Hong Kong, and seeing a completely different culture was so cool. It made me want to visit a lot more countries outside Europe!