A federated learning scheme based on personalized differential privacy and secret sharing

In traditional federated learning, participants train models by sharing model data. However, most existing federated learning frameworks are based on a uniform privacy budget, which cannot meet personalized privacy requirements. Moreover, during the model aggregation process, few federated learning frameworks consider potential malicious attacks or data leakage risks among participants. To address these issues, this paper proposes a federated learning scheme based on personalized differential privacy and secret sharing (PDPSS-FL). This scheme provides personalized differential privacy protection for participants, adding personalized noise to the model to preserve their privacy. Secret sharing techniques are employed during model updates and parameter transmissions to ensure secure model updates in the presence of honest-but-curious servers. Experimental results demonstrate that the proposed scheme generates high-quality models while satisfying personalized privacy protection requirements in a secure environment.