FedPDA: Personalized federated learning based on attribute similarity migration
Xiang Zhou, Qiang Zhi, Ziyang Liu, Dongyi Han, Nan Liu · Information Sciences · 2025
Federated learning enables multiple clients to collaboratively train machine learning models without sharing raw data. Personalized federated learning extends this by tailoring models to each client's local data. To improve privacy, efficiency, and scalability, we propose a personalized federated learning scheme based on attribute similarity migration (Federated learning with Personalized Dynamic Attribute migration, FedPDA). In FedPDA, each client maintains its own model while adaptively transferring knowledge from others based on attribute similarity. To handle data heterogeneity and device diversity, we design SMART (Similarity-guided Model Aggregation with Resource-aware Transfer). By computing multidimensional similarity between clients, SMART enables selective model parameter sharing. To ensure secure knowledge transfer, we incorporate CP-ABE encryption, allowing only qualified clients to decrypt and utilize transferred models under defined attribute policies. This approach achieves a strong balance between model performance and privacy. Extensive experiments on three benchmark datasets show that FedPDA outperforms existing methods, improving accuracy by 5%–11%, reducing communication costs by 18%–35%, and shortening training time by 8%–27% across various scenarios.