面向6G网络的联邦学习资源协作激励机制设计

第6代(6G)移动通信在满足人类智能通信需求的同时, 也给用户数据的安全与隐私保护带来了极大挑战。为此, 基于联邦学习的分布式机器学习架构应运而生。然而, 在联邦的模型训练过程中, 移动设备会产生大量计算和通信开销。自私的移动设备不愿意参与模型训练, 这将降低联邦学习性能。本文基于迭代双边拍卖设计了一种有效的联邦学习资源协作激励机制, 任务计算终端作为卖方, 任务请求终端作为买方, 本地接入点根据买卖双方的出价做出模型训练时延和相应定价决策, 在买卖双方信息非对称情况下最大化联邦学习市场总效用。仿真实验表明, 所提机制具有良好的收敛性, 可显著提高联邦学习的准确率, 同时降低训练损失。

Incentivizing Resource Cooperation for Federated Learning in 6G Networks

Recent advances in 6th Generation (6G) mobile networks can meet the needs of deeper intelligent communication. Meanwhile, it also brings great challenges to the security and privacy preservation of user data. Federated learning is emerging as a distributed learning method to preserve privacy by enabling users to train machine learning models locally and requiring the users to upload only model parameters instead of sending original data to the server. However, the users with mobile devices may be unwilling to participate in federated learning tasks due to considerable overhead for computation and communication, which degrades performance and hinders the application of federated learning. In this study, an iterative double auction is designed based on resource cooperation scheme to incentivize the mobile devices to contribute their resources in federated learning tasks, where model trainers act as sellers and task requesters as buyers. Access point determines optimal training time cost and payment according to the prices offered by the sellers and the buyers. The goal of our incentivizing resource cooperation scheme is to maximize the total utility of federated learning market under information asymmetry. Numerical results show that the proposed scheme can converge to the optimal solution, and also can significantly improve model accuracy and degrade model loss value.