MetisFL

Why we included this project

Organizations working with data that is too sensitive or too scattered to move to one place get a clear payoff from MetisFL: the model trains across many sites while raw examples never leave their origin. It stands out from the federated learning crowd by being built mostly in C++, with the Python API layered on a faster native core, and the authors frame the design around keeping the orchestrator efficient as the number of participating clients grows. Reproducibility is a real concern with a mixed C++/Python stack, so the project ships Docker images and CUDA builds to smooth over the dependency differences between operating systems. That privacy model, where updates rather than data travel, points naturally at security-conscious fields like healthcare and finance. It is research-grade software though: containerized deployment is the intended route, and expect some setup work before training behaves.