Trusted Personalized Federated Learning Based on Differential Privacy

As an emerging machine learning model, federated learning ensures that user data is stored locally while breaking data siloes, which ensures the privacy of training data. However, in practical applications, the training data across different user terminals are often non-independent and identically distributed. Moreover, federated learning does not provide strict privacy protection for users. Attackers can infer data information based on the model parameters or gradients uploaded or downloaded, and there is a great risk of privacy leakage in the process of information transmission. Therefore, we design a personalized federated learning scheme with privacy protection capability. Considering that non-independent and identically distributed data will have a negative impact on model training, we aim to obtain a personalized model for each client by joining transfer learning and knowledge distillation. Since we first propose a personalized federated learning algorithm with privacy protection ability based on the Gaussian mechanism in differential privacy, and then compared with the traditional Laplace mechanism. Experimental results demonstrate that our proposed method achieves better accuracy and improves the privacy protection ability.