基于相似度感知和选择性通信协议的个性化联邦学习框架  

Personalized federated learning framework based on similarity-sense and selective communication protocols

作　　者：李严 庄孟谕 LI Yan;ZHUANG Meng-Yu(School of Management and Economics,University of Electronic Science and Technology of China,Chengdu 611731,China;School of Economics and Management,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]电子科技大学经济与管理学院,成都611731 [2]北京邮电大学经济管理学院,北京100876

出　　处：《四川大学学报(自然科学版)》2025年第2期443-450,共8页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(62172056);中国人工智能学会青年人才托举工程(2022QNRC001)。

摘　　要：提出一种名为相似度感知选择性知识蒸馏(TSKD)的个性化联邦学习框架,旨在解决传统联邦学习框架在通信效率和模型定制方面的局限性.TSKD框架通过设置一个小规模预加载的参考数据集,使本地用户设备能够生成通信凭证并基于此评估其与异构设备网络内其余设备的相似度.根据这个相似度指标,TSKD为本地用户设备分配协作对象并令本地模型与之进行知识共享,进而在保证本地模型个性化的前提下提高模型的性能.在三个真实世界数据集上进行的实验表明,TSKD在各项评估指标上的表现均优于传统的中心化和去中心化学习方法,且能够在资源受限的环境中高效地实现知识共享,提升模型的准确性和个性化程度.With the advancement of edge computing and wearable technologies,a notable increase in the deployment of deep neural networks(DNNs)on endpoint devices has been observed.A personalized federated learning framework named Similarity-Sense Selective Knowledge Distillation(TSKD)is proposed to address the limitations of traditional federated learning frameworks in terms of communication efficiency and model customization.The framework utilizes a small-scale preloaded reference dataset,enabling local user devices to generate communication credentials and assess their similarity with other devices in a heterogeneous network.Based on this similarity,devices selectively share knowledge with the most similar devices,thereby enhancing the performance of their local models.Experiments conducted on three real-world datasets demonstrate that TSKD outperforms traditional centralized and decentralized learning methods across various evaluation metrics.Furthermore,TSKD efficiently facilitates knowledge sharing in resource-constrained environments,improving both model accuracy and personalization.

关 键 词：联邦学习 个性化分析 知识蒸馏 数据异质 异构问题 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]