EuProGigant: A decentralized Federated Learning Approach based on Compute-to-Data and Gaia-X
Roman Gehrer, Stefan Dumss, Fabian Gast, Willi Wünschel, Frederic Schwill, Mateo Šoša, Shiyang Zhou, Gerald H. Ristow, Tatevik Gharagyozyan, Clemens Heistracher, Manfred Grafinger, Matthias Weigold · Procedia CIRP · 2024
Machine learning normally requires a considerable amount of data for model training, limiting the field of usage especially for small manufacturing companies due to their lack of machine data. Federated learning provides an opportunity to enlarge the data basis for model training without the need to directly share the machine data with other participants, addressing concerns regarding to privacy, intellectual property and potential reverse engineering of proprietary process information through competitors. Previous research focused mainly on federated learning models, mostly managed and orchestrated by some kind of centralized authority. The presented approach shows a more decentralized, self-sovereign concept of federated learning for the manufacturing industry, expanding its applicability to a broader range of participants. It combines existing solutions for machine learning and Compute-to-Data by Ocean Protocol with the concept of dataspaces as defined by Gaia-X. The methodology is demonstrated through an use case derived from industry demands involving several CNC milling machines.