Homomorphic encryption(HE) enabled federated learning

Abstract

In this report, to maximise data privacy, we conducted Federated Learning algorithm with Homomorphic Encryption. The project was done in stages. Initially, federated learning was done without applying homomorphic encryption. Homomorphic encryption was applied in a progressive manner at a later stage and its performance was thoroughly studied. Additionally, the existing projects incorporating homomorphic encryption was studied to further improve our project. Various parameters pertaining to homomorphic encryption were also explored to observe the key features and necessary trade-offs. Based on the testing results, it was discovered that the prediction accuracy was relatively higher for the ML models generated from the averaged weights within the federated network. For future work, different datasets will be used to further confirm this finding.