Personalized Federated Learning with Partial Model Sharing and Client-Customized Aggregation

Personalized federated learning (PFL) addresses the limitations of traditional federated learning (FL) in statistically heterogeneous scenarios, where diverse client data distributions reduce model applicability. However, many PFL methods compromise client privacy or depend on external client information. This paper proposes a novel PFL method that enhances privacy and autonomy by enabling each client to share only partial model components and adaptively optimize local aggregation weights to align with its local objective. This approach minimizes reliance on other clients' information while ensuring robust personalization. Comparative experiments on three real-world datasets against seven baseline methods demonstrate that our method achieves higher or comparable accuracies, validating its effectiveness in non-IID settings.