Abstract: To address privacy leakage and performance degradation in federated learning with heterogeneous data, we propose a verifiable privacy-preserving personalized federated learning scheme. In the scheme, the privacy of users is guaranteed through homomorphic encryption. Personalized model customization is enabled by calculating similarities over ciphertexts. Based on the ring learning with errors problem, users can verify the correctness of personalized updates. Theoretical and experimental analysis shows that the proposed scheme effectively preserves user privacy, ensuring that neither the server nor the user can access others’ local or personalized updates. Furthermore, the additional computational and communication overhead incurred by privacy preservation remains within acceptable limits. Experimental results on two public datasets show that the proposed scheme achieves higher accuracy than federated averaging and other personalized schemes under both independently and non-independently distributed data settings.

Key words: privacy preserving, federated learning, homomorphic encryption, personalized, verifiable