FedUNet: A lightweight additive U-Net module for federated learning with heterogeneous models

Federated learning (FL) enables decentralized model training without sharing data, but most existing methods assume homogeneous model architectures across clients, limiting applicability in heterogeneous real-world environments. We propose FedUNet , a lightweight and architecture-agnostic FL framework that introduces an additive U-Net-inspired module attached to each client’s backbone. By sharing only the compact bottleneck of the U-Net, FedUNet enables efficient knowledge transfer without architectural alignment, shared data, or costly distillation. The encoder–decoder structure with skip connections integrates multi-scale features into a shared latent space, promoting client-invariant representation alignment across heterogeneous models. Experimental results on CIFAR-10 and CIFAR-100 with VGG and ResNet backbones show that FedUNet achieves high accuracy with significantly reduced communication and computational overhead. Notably, a full U-Net attains 93.92% accuracy while requiring only 0.89 MB of communication per round, highlighting its practicality for resource-constrained FL systems.