FedDK: Toward Adaptive Personalized Federated Learning via Decoupled Knowledge Transfer Under Heterogeneous Environments

  • Han, Seungjin; 
  • Lee, Changwoo
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Federated learning (FL) enables collaborative training without sharing raw data; however, its performance typically degrades in heterogeneous client environments due to non-independent and identically distributed (non-IID) data, model disparity, and conflicting objectives between personalization and global generalization. To address this issue, we propose a novel federated decoupled knowledge (FedDK), a personalized FL framework that introduces a dual-head architecture with explicit role separation. FedDK factorizes the output layer into a balanced head that learns from both private supervision and globally distilled knowledge to preserve client-specific adaptation, and a global head that is optimized exclusively with server-aggregated soft targets on proxy data to maintain client-agnostic global knowledge. This decoupled design mitigates optimization interference between local and global objectives while avoiding full parameter exchange via proxy-based knowledge transfer. Extensive experiments on EMNIST, CIFAR-100, and SVHN under strong, intermediate, and weak non-IID settings and cross-domain proxy scenarios, demonstrate that FedDK exhibits high and consistent performance across diverse heterogeneous environments, including domains with very limited proxy data. Overall, FedDK provides a lightweight and practical framework for balancing personalization and global consistency in heterogeneous federated learning.

키워드

distillation; dual head; Federated learning; student model; teacher model
제목
FedDK: Toward Adaptive Personalized Federated Learning via Decoupled Knowledge Transfer Under Heterogeneous Environments
저자
Han, Seungjin; Lee, Changwoo
DOI
10.1109/ACCESS.2026.3703950
발행일
2026
유형
Article
저널명
IEEE Access
권
14
페이지
91468 ~ 91478