Secure architectures implementing trusted coalitions for blockchained distributed learning (TCLearn)

In some fields (e.g. the medical field), centralizing a large amount of data is hard to perform because the data owners want to keep it confidential. This leads to the development of distributed methods that allows to train a machine learning model without centralizing the data. This thesis focus on the improvement of one of these methods called TCLearn. The TCLearn architectures allow data owners to jointly train a machine learning model without trusting each other. Indeed, each member's proposal of model improvement will be evaluated by the other members through a protocol that we call the Model Evaluation Protocol (MEP). This protocol is used to decide if the model proposal will be chosen as the new model or will be rejected. The purpose of this protocol is to reject model proposal that result from wrong training or that are designed for model degradation. But we also want to avoid a too restrictive MEP that will reject too much model proposal resulting from good training. A first version of the MEP was already implemented, we focus the work of this thesis on the improvement of this version based on statistical tools.