Federated Learning System for Imbalanced Data Using Secure Computation

Abstract

We constructed a gradient data sharing system to perform federated learning on unbalanced data held by each user of the system. First, we considered issuing temporary keys as key delivery among many users to reduce the cost of key management and to prevent collusion between the computation server and the model administrator. Furthermore, we weighted the gradients sent by the participants while keeping the data size and imbalance ratios secret to compute gradients that would be effective for AI learning. We proposed a system that can effectively operate a federated learning system.