Federated Learning with Trust Evaluation for Industrial Applications

We propose and investigate a novel Reputation and Trust-based technique incorporation for Federated Learning (FL) industrial applications to address possible anomalous local data caused by malicious attacks or other factors that can deteriorate the produced Machine Learning (ML) model performance. We evaluate our technique against the centralized Deep Neural Network- and Random Forest-based topologies and also against conventional FL structures on the available financial industry datasets. Our study results demonstrate that our techniques allows detecting anomalous local data models and also improving the produced model effectiveness, especially in the cases of practical importance.