Adaptive Homomorphic Encryption for Federated Learning

Abstract

Federated learning (FL) offers a promising paradigm for training machine learning models on decentralized data without direct data sharing. However, the traditional approach of using static homomorphic encryption (HE) schemes for FL introduces significant overheads, particularly when combined with the iterative nature of the training process. This paper proposes a novel adaptive homomorphic encryption scheme designed specifically for federated learning. The core idea revolves around dynamically adjusting the ciphertext structure of the HE scheme during training, optimizing for computational efficiency and security. We introduce a system where the encryption key evolves based on the data distribution and the learning algorithm employed. This adaptive approach mitigates the inefficiencies of static HE schemes and represents a significant step towards realizing the full potential of HE in FL. The proposed scheme is theoretically sound and provides a framework for future research in this area.