Secure and efficient multi-key aggregation for federated learning

Abstract

Federated learning (FL) is a distributed machine learning framework that aims to provide privacy for local datasets while learning a global machine learning model. However, the updates exchanged in FL may indirectly reveal information about the local training datasets. To protect the confidentiality of updates, various solutions using cryptographic schemes such as secret sharing, differential privacy, and homomorphic encryption have been developed. The solutions using secret sharing often have high communication costs, while those using differential privacy require a trade-off between accuracy and privacy. Homomorphic encryption has been used to address these challenges to reduce communication costs and provide provable security. However, existing solutions based on homomorphic encryption require clients to use the same public-private key pair, which may lead to local updates disclosure. To address the existing challenges, we propose a secure and efficient multi-key aggregation protocol (MKAgg) that utilizes homomorphic encryption. The protocol allows clients to drop out at any point during the process. We construct MKAgg based on a two-server model and adopt a proxy re-encryption scheme with additively homomorphic properties to implement secure and efficient ciphertext transformation and calculation. We provide security proof to demonstrate that our MKAgg protocol meets the required security standards. Furthermore, we perform an efficiency analysis of MKAgg and evaluate its performance on various datasets. The results affirm that MKAgg is both effective and efficient for aggregation in the multi-key setting. We then apply MKAgg in FL and develop a multi-key privacy-preserving neural network scheme called MKPNFL. We analyze the security of MKPNFL and conduct tests using real-world datasets. The results demonstrate that MKPNFL is secure and practical for real-world applications.