通信学报2026，Vol.47Issue(7)：123-136,14.DOI:10.11959/j.issn.1000-436x.TXXB260116
基于特征投影与自适应增强的联邦蒸馏方法
Federated distillation method based on feature projection and adaptive enhancement
摘要
Abstract
To address the problem that existing federated distillation methods struggle to handle inconsistent representa-tion spaces and uneven feature distributions across heterogeneous models,a federated distillation method based on fea-ture projection and adaptive enhancement was proposed,achieving efficient knowledge fusion across heterogeneous cli-ent models.On the server side,client outputs were integrated via distillation.On the client side,lightweight multi-exit branches were employed to project intermediate features into an aligned logits space to overcome feature alignment bottlenecks.Promising performance was demonstrated on benchmark datasets,with reduced communication rounds and data transmission as well as enhanced system robustness,thus providing an effective new solution for heterogeneous model knowledge fusion in federated learning.关键词
联邦学习/知识蒸馏/异构模型/特征投影Key words
federated learning/knowledge distillation/heterogeneous model/feature projection分类
信息技术与安全科学引用本文复制引用
陈宁江,章德华..基于特征投影与自适应增强的联邦蒸馏方法[J].通信学报,2026,47(7):123-136,14.基金项目
国家自然科学基金资助项目(No.62162003) （No.62162003）
广西重点研发计划资助项目(No.桂科AB25069258,No.桂科AB25069130)The National Natural Science Foundation of China(No.62162003),The Guangxi Key Research and Develop-ment Program(No.GuikeAB25069258,No.GuikeAB25069130) （No.桂科AB25069258,No.桂科AB25069130）
