MULTI-KEY FULLY HOMOMORPHIC ENCRYPTION FOR PRIVACY-PRESERVATION WITHIN FEDERATED LEARNING ENVIRONMENTS

Abstract

Despite the need for data from multiple sources in machine learning, privacy constraints limit data sharing. Federated Learning (FL) addresses this by allowing clients to share locally trained model parameters without disclosing sensitive data, however, recent research highlights data leakage risks. This paper investigates multi-key fully homomorphic encryption, specifically MK-CKKS, to enhance data privacy in FL. The study demonstrates MK-CKKS’s effectiveness in protecting model parameters transmission and preventing external access to private information. Nonetheless, precautions are needed during decryption, as vulnerabilities may allow the aggregator server and external adversaries to infer personal data from shared partial descriptions, impacting the client’s data privacy and security.

Disclosure statement

No potential conflict of interest was reported by the author(s).