With the fourth industrial revolution, the construction of the Internet of Things (IoT) has developed vigorously, and machine learning is also widely used in IoT management and data processing. Given the existence of massive distributed and private datasets generated by a large number of IoT devices, centralized machine learning is unsatisfactory. Therefore, federated learning (FL), as a distributed learning method, becomes a promising solution. In FL, clients can train models by transferring model parameters to the aggregation server while keeping private data locally. However, FL still relies on a central server, which has questionable reliability. The single point of failure and limited communication resources also hinder the application of FL in the IoT. In this paper, we propose a voting consensus based decentralized federated learning method (VCDFL) by incorporating the leader-candidate-follower hierarchical management method and the consensus based leader election mechanism to solve the single point of failure and exclude outlier models for accelerating convergence during aggregation. Then, we propose a joint decision method to exchange decision information rather than model transfer between clients to further protect privacy and reduce communication overhead while ensuring accuracy. Furthermore, we mathematically derive the probability of successfully electing a leader, the communication efficiency and the joint decision accuracy. We conduct our method in an image recognition scenario. The results show that our joint decision mechanism promotes the accuracy of both system and local decision-making. Meanwhile, the proposed scheme greatly reduces communication costs compared to benchmark learning methods.

In recent years, due to the increasing concern about data privacy security, federated learning, whose clients only synchronize the model rather than the personal data, has developed rapidly. However, the traditional federated learning system still has a high dependence on the central server, an unguaranteed enthusiasm of clients and reliability of the central server, and extremely high consumption of communication resources. Therefore, we propose Clustered Hierarchical Distributed Federated Learning to solve the above problems. We motivate the participation of clients by clustering and solve the dependence on the central server through distributed architecture. We apply a hierarchical segmented gossip protocol and feedback mechanism for in-cluster model exchange and gossip protocol for communication between clusters to make full use of bandwidth and have good training convergence. Experimental results demonstrate that our method has better performance with less communication resource consumption.