Differential Privacy for Federated Learning via Homomorphic Encryption

Abstract

This paper presents a novel approach to enhancing data privacy in federated learning by integrating differential privacy with homomorphic encryption. Federated learning, a distributed machine learning paradigm, allows training models across multiple devices without directly sharing raw data. However, this approach still exposes participants to potential privacy risks. Our method leverages homomorphic encryption to enable computations on encrypted data, mitigating the risk of direct data exposure. Subsequently, we apply differential privacy techniques to the aggregated model updates, providing a rigorous guarantee against individual data disclosure. The core claim is that this combined strategy offers a robust solution for preserving data privacy in federated learning settings. We outline the key mechanisms involved, demonstrating how homomorphic encryption is used to perform computations on encrypted data and how differential privacy is applied to the resulting aggregated model updates. The proposed framework addresses the unique challenges posed by federated learning, offering a significant improvement in privacy protection.