Abstract: Federated Learning (FL) is driven by multi-party data participation,and it maximizes the value of the data itself through data encryption interaction.In recent years,FL has attracted extensive attention from researchers from all walks of life and gradually moved from basic theoretical research to practical applications,which provides new technologies for further exploiting the value of data for enterprises.Based on the definition and classification of FL,a comprehensive analysis and summary of the research progress of related technologies at home and abroad was conducted,including privacy protection,communication efficiency,heterogeneity,and incentive mechanisms.The current application platforms and frameworks of FL were introduced,and the application frameworks of FL was proposed in the fields of intelligent manufacturing,medical treatment and education.Combined with the deficiencies of FL in some key open issues,its future development trends and directions were summarized for providing a reference for the theoretical research and applications of FL.

Key words: federated learning, privacy protection, communication efficiency, heterogeneity, incentive mechanism

摘要： 联邦学习(FL)以多方数据参与为驱动,通过数据加密交互实现数据自身价值的最大化,近年来受到各界研究学者的广泛关注与研究,逐步从基础理论研究走向实际应用,为企业进一步发挥数据价值提供了新技术。在阐述联邦学习定义及分类的基础上,首先对其隐私保护、通信效率、异构性、激励机制等相关技术的国内外研究进展展开了较为全面的分析和总结;其次介绍了当前联邦学习已有的应用平台和框架,并提出了联邦学习在智能制造、医疗、教育等领域的应用框架;最后,结合联邦学习在一些关键的开放性问题上的不足,对其未来发展趋势和方向进行了总结与展望,旨在为联邦学习的理论研究及应用落地提供参考。

关键词: 联邦学习, 隐私保护, 通信效率, 异构性, 激励机制