PFL-IDGAN:Personalized Federated Learning Framework Based on Interactive Dual Generative Adversarial Networks

Zhigang Wang , Yan Yang , Xiaochi Hou , Junfeng Zhao ·

YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.3233/faia250942

Federated learning (FL) enables collaborative model training without direct data exchange, promoting privacy-preserving data utilization. To address performance degradation caused by non-independent and identically distributed (non-IID) data, Personalized Federated Learning (PFL) allows each client to learn a model tailored to its local distribution. However, real-world personalized scenarios often involve not only data heterogeneity but also model heterogeneity across clients. Existing PFL methods struggle under the coexistence of both, as parameter aggregation requires identical model structures, while knowledge distillation often relies on shared public data. To tackle these challenges, we propose a novel PFL framework called Personalized Federated Learning based on Interactive Dual Generative Adversarial Networks (PFL-IDGAN). This framework leverages Generative Adversarial Networks (GANs) to augment local datasets, effectively mitigating label discrepancies and non-iid. data issues across clients. Moreover, it introduces a dual adversarial learning mechanism that enables fine-grained knowledge transfer and collaboration across clients, while supporting heterogeneous model architectures. Extensive experiments demonstrate that the proposed PFL-IDGAN framework significantly outperforms existing baseline methods, particularly in settings with pronounced disparities in client models and data distributions.