Fig. 2
From: Splitting chemical structure data sets for federated privacy-preserving machine learning
Federated Machine learning scheme in MELLODDY. a MELLODDY uses a feed-forward neural network, which is divided in a trunk and head part. The trunk is jointly trained by all partners, using secure aggregation of parameter updates at each iteration. The head part is individual to each pharma partner, an thus allows each partner modelling their own assays without mapping to the other partner’s assays. The common trunk enables transfer learning between the tasks of different partners. b The technical architecture of MELLODDY. Each partner maintains a node on a cloud platform. An additional model dispatcher node aggregates the trunk updates, but is prevented by the secure aggregation protocol from accessing these updates itself