Enhancing Privacy and Efficiency in Federated Learning Through Hybrid Homomorphic Encryption

Abstract

Federated Learning (FL) allows the training of models over distributed data sources without compromising the privacy of users in different client devices. Nonetheless, encryption mechanisms, including Homomorphic Encryption (HE) and symmetric encryption, such as the Advanced Encryption Standard (AES), enhance security but usually come at the cost of computational expense, affecting the speed and scale of the model. To tackle these issues, we propose a hybrid HE model which integrates the CKKS method of HE with AES encryption to ensure privacy and computational performance. Our experiments with an agricultural crop production dataset indicate that the proposed model is considerably more efficient regarding predictive performance and training time than the normal encryption model. The hybrid method also outperforms security and performance resource-constrained systems compared to systems that use only AES or HE. This implies that a hybrid strategy can be used in FL with the objective of achieving both security and efficiency.