Federated learning, as an innovative machine learning technique, not only effectively protects user data privacy and enhances data efficiency but also supports collaboration among multiple participants, demonstrating its extensive application potential. Nevertheless, federated learning still faces some limitations in practical applications. Specifically, when client data are non-independent and identically distributed (Non-IID), the model’s generalization performance and convergence speed are significantly affected. Additionally, the gradients obtained by users during local model training may be reconstructed by attackers to reveal the complete training data. Addressing these issues of poor model generalization in Non-IID data scenarios and potential privacy leakage during gradient uploads in federated learning, this paper proposes a personalized federated learning algorithm, SS-PFL, based on shuffled differential privacy. The algorithm initially adopts a bi-level optimization approach to train personalized models locally on clients, optimizing local model performance. Secondly, it introduces a shuffled server to ensure secure aggregation of gradient parameters within a unified privacy budget scenario. Finally, experiments are conducted to verify whether SS-PFL can effectively resist reconstruction attacks while improving accuracy.