Personalized federated learning with feature complementarity and temporal regularization

Abstract Personalized federated learning aims to deliver customized models for clients with heterogeneous data distributions while preserving data privacy. In this work, we propose personalized federated learning with feature complementarity and temporal regularization (pFedCT), a novel PFL framework where each client adopts a broad learning system (BLS) as the local model. Indeed, the existing BLS-based FL methods have been successful in enhancing communication efficiency, privacy preservation, and Byzantine robustness, but they often fail to address the critical issue of personalization. To achieve effective personalization, pFedCT introduces two key components. First, we extract feature node matrices from each client’s BLS model, which capture the internal feature representations of local data. By comparing the principal subspaces of these matrices, we assess the complementarity between clients’ learned representations. This enables each client to collaborate with others who provide diverse and informative knowledge. Second, we incorporate a temporal regularization mechanism to stabilize the collaboration structure across communication rounds, mitigating performance degradation caused by fluctuating local updates and data shifts. Extensive experiments on Modified National Institute of Standards and Technology (MNIST) and Fashion-MNIST datasets under both independent and identically distributed (IID) and non-IID settings demonstrate that pFedCT consistently outperforms existing BLS-based baselines and achieves competitive accuracy compared to deep neural network-based methods, while maintaining superior efficiency in communication and computation.