희소 표현 기반 개인화 연합학습 기법에 관한 연구 = A Study of Personalized Federated Learning Based on Sparse Representations

다국어 초록 (Multilingual Abstract)

Federated Learning (FL) enables collaborative training in distributed environments, but suffers from performance degradation and inter-client disparity under Non-IID data. This paper proposes a personalized federated learning method that combines a sparse representation–based global model with a lightweight personalization layer. The proposed approach maintains computational efficiency and training stability via a k-sparse constraint, while enabling local adaptation without additional communication overhead.
Experiments on CIFAR-10 and CIFAR-100 under various Non-IID settings (α = 0.1, 0.3, 0.5, 1.0) and client scales (10, 20, 50) show that the proposed method preserves global accuracy while reducing inter-client variance. In particular, it significantly improves the performance of low-performing clients in highly heterogeneous environments, demonstrating robust and scalable performance across diverse settings.