Machine Learning for All: A More Robust Federated Learning Framework

Abstract

Machine learning and especially deep learning are appropriate for solving multiple problems in various domains. Training such models though, demands significant processing power and requires large data-sets. Federated learning is an approach that merely solves these problems, as multiple users constitute a distributed network and each one of them trains a model locally with his data. This network can cumulatively sum up significant processing power to conduct training efficiently, while it is easier to preserve privacy, as data does not leave its owner. Nevertheless, it has been proven that federated learning also faces privacy and integrity issues. In this paper a general enhanced federated learning framework is presented. Users may provide data or the required processing power or participate just in order to train their models. Homomorphic encryption algorithms are employed to enable model training on encrypted data. Blockchain technology is used as smart contracts coordinate the work-flow and the commitments made between all participating nodes, while at the same time, tokens exchanges between nodes provide the required incentives for users to participate in the scheme and to act legitimately.