KApFL: Kolmogorov-Arnold Networks for Personalized Federated Learning
Rou Zhou, Yuling Chen, Hui Dou · IEEE Transactions on Consumer Electronics · 2025
The heterogeneity of client data distributions in federated learning makes a single global model likely to perform poorly on some devices. Recently, Personalized Federated Learning (PFL) has attracted much attention due to its ability to train models tailored to each client’s unique needs. However, in real deployments, client–server objectives are not fully aligned, and client participation is highly dynamic: arbitrary joining and dropout during training not only reduce global accuracy but also cause fairness degradation across clients, increased model drift, and unstable personalized performance. Most existing PFL methods, while effective at mitigating inter-client data heterogeneity, fail to explicitly address such dynamic participation, leaving models vulnerable to instability and degraded robustness in practice. To address these issues, we propose a novel PFL method, named KApFL, which introduces a probabilistic extension JacobiKAN to Kolmogorov–Arnold Networks (KAN) based on Jacobi polynomials and learns personalized model parameters for it using user descriptors collected from clients. KApFL provides a highly generalized, plug-and-play personalized training framework that improves robustness when clients join or leave at will. superiority of KApFL over ten state-of-the-art methods in performance, scalability, and stability, achieving up to a 7.74% improvement over the best-performing baseline.