uFedMba: Unforgotten Presonalized Federated Learning with Memory Bank for Adaptively Aggregated Layers
Personalized federated learning (PFL) addresses the challenge of data heterogeneity across clients. However, existing efforts often struggle to balance model personalization and generalization under Non-IID data scenarios. This paper proposes uFedMba, a novel PFL framework that decouples neural network parameters into global base-layer parameters and client-specific personalized-layer parameters. On the client side, uFedMba adds a penalty term with base-layer parameters into the local loss function to prevent overfitting to local data and integrates the historical model into personalized-layer parameters for accelerating convergence. The server employs layer-wise aggregation based on gradient alignment to adaptively aggregate personalized layers, enhancing compatibility across heterogeneous clients. Extensive experiments demonstrate that the uFedMba achieves state-of-the-art results on four image classification datasets.