This research paper introduces Fed-CO2, a universal Federated Learning (FL) framework designed to handle both label distribution skew and feature skew. It uses a cooperation mechanism between online and offline models. The online model learns shared knowledge among clients, while the offline model learns the specialized knowledge of each client. An intra-client knowledge transfer mechanism is designed to reinforce mutual learning between these models. The paper shows that Fed-CO2 outperforms existing personalized federated learning algorithms in handling label distribution skew and feature skew.
Publication date: 21 Dec 2023
Project Page: https://arxiv.org/abs/2312.13923v1
Paper: https://arxiv.org/pdf/2312.13923
