PBPAFL: A Federated Learning Framework with Hybrid Privacy Protection for Sensitive Data

Abstract

Due to the difficulties of exchanging data securely, data silos have become a critical issue in the era of big data. Federated learning provides an advantageous approach by enabling data holders to train a model collaboratively without sharing local data. However, multiple known inference attacks have made it impossible for a purely federated learning approach to protect privacy well enough. We present a PBPAFL algorithm that combines differential privacy with homomorphic encryption based on federated learning with an assessment module that enables the privacy budget parameters to be flexible in response to varying training requirements. The models trained using our proposed PBPAFL algorithm are capable of preventing inference assaults without a severe loss of precision. To demonstrate the efficacy of our proposed framework, we employ the PBPAFL algorithm to train a collection of face image-sensitive data. The experimental results show that our approach can improve the privacy protection of the model while maintaining precision.