【Author】
Kim, You Jun; Hong, Choong Seon
【Source】2019 20TH ASIA-PACIFIC NETWORK OPERATIONS AND MANAGEMENT SYMPOSIUM (APNOMS)
【Abstract】Federated learning (FL) is a decentralized learning method that deviated from the conventional centralized learning. The FL progresses learning locally on each device and gradually improves the learning model through interaction with the central server. However, it can cause network overload because of limited communication bandwidth and the participation of a huge number of users. One of the ways to minimize the network load is for the model to converge rapidly and stably with target learning accuracy. In this paper, we propose blockchain based federated learning scenario. Blockchain can efficiently induce users to participate in learning and can separate each participating user as a 'node'. In addition, it can be pursued the integrity, stability, and so on. We consider two types of weights to choose the subset of clients for updating the global model. First, we consider the weight based on local learning accuracy of each client. Second, we consider the weight based on participation frequency of each client. We choose two key performance indicators, learning speed and standard deviation, to compare the performance of our proposed scheme with existing schemes. The simulation results show that our proposed scheme achieves higher stability along with fast convergence time for targeted accuracy compared to others.
【Keywords】Federated Learning; Blockchain; Node Selection; Weighting scheme
【标题】用于提高联邦学习性能的基于区块链的节点感知动态加权方法
【摘要】联邦学习（FL）是一种偏离传统集中学习的去中心化学习方法。 FL 在每个设备上进行本地学习，并通过与中央服务器的交互逐步改进学习模型。但是，由于通信带宽有限和大量用户的参与，它会导致网络过载。最小化网络负载的方法之一是使模型快速稳定地收敛，并具有目标学习精度。在本文中，我们提出了基于区块链的联邦学习场景。区块链可以有效地诱导用户参与学习，并且可以将每个参与用户分离为一个“节点”。此外，还可以追求完整性、稳定性等。我们考虑两种类型的权重来选择客户端子集来更新全局模型。首先，我们根据每个客户端的本地学习精度来考虑权重。其次，我们根据每个客户的参与频率考虑权重。我们选择两个关键性能指标，学习速度和标准差，来比较我们提出的方案与现有方案的性能。仿真结果表明，与其他方案相比，我们提出的方案在目标精度方面实现了更高的稳定性和更快的收敛时间。
评论