Decentralized Federated Learning via Mutual Knowledge Transfer

In this article, we investigate the problem of decentralized federated learning (DFL) in Internet Things (IoT) systems, where a number IoT clients train models collectively for common task without sharing their private training data absence central server. Most existing DFL schemes are composed two alternating steps, i.e., model updating and averaging. However, averaging parameters directly to fuse different at local suffers from client-drift, especially when heterogeneous across clients. This leads slow convergence degraded performance. As possible solution, propose via mutual knowledge transfer (Def-KT) algorithm, by transferring learned each other. Our experiments on MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100 sets reveal that proposed Def-KT algorithm significantly outperforms baseline methods with averaging, Combo FullAvg, not independent identically distributed (non-IID)