Trust-Model-Based Method for Protecting Global Models in Federated Learning Systems

Abstract

The problem of ensuring the security of a global computational model in federated learning systems is considered. A method is proposed that is based on data verification using a trusted group of nodes and ensures that only correct updates are taken into account during the global model aggregation process. It is experimentally demonstrated that the developed method ensures accurate identification and isolation of adversaries implementing label-flipping and noise-injection threats.