APPFed: A Hybrid Privacy-Preserving Framework for Federated Learning over Sensitive Data

Abstract

In the era of Big Data, data silos have become a pressing problem due to the difficulty of secure data sharing. Federated learning provides a favorable solution by allowing data holders to collaborate in training a model without sharing local data. However, several existing inference attacks have led to the fact that a pure federated learning methodology is incapable of providing sufficient privacy protection. We propose an APPFed algorithm that combines differential privacy and homomorphic encryption based on federated learning, where exists an evaluation module that enables the privacy budget parameters to be adaptive according to different needs during the training. Trained with our proposed APPFed algorithm, the models are enabled to prevent inference attacks without drastic accuracy depletion. To verify the effectiveness of our proposed algorithm, we use the APPFed algorithm to train a set of sensitive data containing face images. The experimental results show that our approach can enhance privacy protection while balancing model accuracy.