FedSnM: P2P 네트워크에서 효율적인 통신을 위한 Score-and-Model 방식을 활용한 연합학습.

Federated Learning is a machine learning strategy that is communication-efficient and privacy-preserved. However, Federated Learning in a P2P network environment to provide high scalability requires a larger amount of communications than a server-client. Therefore, we propose FedSnM that Federated Learning strategy using Score-and-Model, which sends the Score derived from the Federated Learning training process in the preceding Push Gossip process and receives the Model trained from the data held by each client in the Pull Gossip process. FedSnM reduces the total number of communications by 88.5% and improves the accuracy by 4.26% compared with FedAvg. In addition, FedSnM improves the bottleneck by reducing the communication cost by 68.3% based on the client who performed the most communication compared with FedPSO. In the future, FedSnM can evolve into the learning strategy in an asynchronous network that has different communication speeds through improved Gossip Protocol. [ABSTRACT FROM AUTHOR]