Skip to main navigation Skip to search Skip to main content

Personalized Federated Learning Via Dual-Decoupled Prototype Calibration

Research output: Book Chapters | Papers in Conference Proceedings › Conference paper (refereed) › Referred Conference Paper › peer-review

Abstract

Federated learning has garnered extensive attention in the realm of Connected and Autonomous Vehicles for its ability to coordinate multi-vehicle collaborative training while preserving data privacy. Given the communication uncertainties and data heterogeneity inherent in moving vehicles, prototypebased personalized federated learning emerges as a practical paradigm for deployment. However, under Non-IID scenarios, existing methods often suffer from performance degradation due to coupled optimization. The joint training of the feature extractor and classifier on clients can induce optimization conflicts, and the passive average aggregation of local prototypes on the server leads to prototype collapse. To address these issues, we propose FedDPC, a Federated Dual-Decoupled Prototype Calibration framework. A Global Prototype Learner is introduced on server with Orthogonal Prototype Contrastive Learning to construct representative and orthogonal global prototypes. And a Top-Down Dual Calibration strategy is designed on clients to decouple the optimization process of the feature extractor and classifier. By leveraging Virtual Prototype Replay to calibrate the classifier, FedDPC effectively mitigates the optimization conflicts between the biased classifier and global prototypes. Extensive experiments demonstrate that FedDPC outperforms existing methods in various data heterogeneity scenarios.
Original languageEnglish
Title of host publication2026 15th Asian Control Conference (ASCC): Proceedings
PublisherIEEE
Pages188-193
Number of pages6
ISBN (Electronic)9798319521071
DOIs
Publication statusPublished - Jun 2026
Event2026 15th Asian Control Conference (ASCC) - Nusa Dua, Bali, Indonesia
Duration: 17 Jun 2026 → 21 Jun 2026

Publication series

NameAsian Control Conference
PublisherIEEE
ISSN (Print)2770-8365
ISSN (Electronic)2770-8373

Conference

Conference2026 15th Asian Control Conference (ASCC)
Abbreviated titleASCC 2026
Country/TerritoryIndonesia
CityBali
Period17/06/26 → 21/06/26

Bibliographical note

Publisher Copyright:
© 2026 IEEE.

Funding

This work was supported by the National Natural Science Foundation of China through Grant Nos. 62306071, U22B2046, and U2541220.

Keywords

  • federated learning
  • orthogonal prototype contrastive learning
  • connected and autonomous vehicles

Fingerprint

Dive into the research topics of 'Personalized Federated Learning Via Dual-Decoupled Prototype Calibration'. Together they form a unique fingerprint.

Cite this