CKKS and Local Differential Privacy based Model Parameters Encryption for Personalized Federated Learning

Federated learning achieves a certain degree of privacy protection by only uploading model parameters instead of the clients’ raw data. However, attackers can still analyze model parameters to obtain sensitive data, and previous studies failed to fully balance the implementation of privacy protection and personalized requirements in federated learning. Therefore, we propose a Cheon-Kim-Kim-Song (CKKS) and Local Differential Privacy based Model Parameters Encryption for Personalized Federated Learning scheme to solve this problem, called CLDPMPE, which consists of two parts: model parameters privacy protection and model personalization. Our scheme ensures the security of model parameters by leveraging the CKKS encryption algorithm in conjunction with local differential privacy techniques. Furthermore, it facilitates model personalization through the application of meta-learning and dynamic clustering methodologies. Since then, while ensuring privacy and security, maintaining the model’s generalization and personalization capabilities. Formal security analysis confirms that CLDP-MPE is secure and comprehensive experiments demonstrate that it achieves a high level of personalized testing precision (averaging 89.58%) while protecting privacy.