FLHyHE: Federated Learning with Hybrid Homomorphic Encryption
Abstract
Federated Learning (FL) enables distributed and collaborative training while keeping sensitive data on clients' devices. This approach is considered more privacy-preserving since only model updates are shared with an aggregator. However, several studies have shown that these updates can still leak private information, making them vulnerable to various attacks. Privacy-preserving techniques, such as Differential Privacy and Homomorphic Encryption, can be combined with FL to enhance security. However, these methods often introduce a loss in accuracy and increase computational and communication overhead on the network. Hybrid Homomorphic Encryption (HHE) is a relatively recent type of cryptographic primitive that has been experimentally applied to FL to mitigate client-side overhead while preserving privacy. In this work, we propose a general construction for FL systems employing HHE, denoted FLHyHE, and analyse its security. We instantiate this construction on top of the Flower framework using the PASTA/FV HHE scheme and evaluate it experimentally. Experiments conducted on two well-known datasets, MNIST and CIFAR-10, using two convolutional neural network models show that model accuracy is mainly preserved for MNIST (≈2% drop), while performance on CIFAR-10 is noticeably affected by the use of HHE (≈20% drop). On the other hand, it is clear that a solution of this kind is significantly more efficient for clients in terms of computational and communication costs compared to solutions relying solely on Homomorphic Encryption, since it uses Symmetric Encryption as the main encryption tool. Especially as models get more complex and consist of more parameters.