"Secure Model Aggregation Techniques for Encrypted Federated Learning"

Abstract

Secure model aggregation techniques play a crucial role in advancing the field of encrypted federated learning (EFL), where preserving data privacy is paramount. In EFL, multiple clients collaboratively train a machine learning model without sharing raw data by encrypting their updates. However, aggregating these encrypted model updates while maintaining security and efficiency remains a challenge. This abstract explores various secure aggregation techniques tailored for EFL, focusing on cryptographic protocols like secure multi-party computation (MPC) and homomorphic encryption (HE). These techniques ensure that the server can aggregate model updates without decrypting individual contributions, thereby safeguarding client data privacy. We discuss the advantages and limitations of each approach, highlighting their applicability in different scenarios. Additionally, we survey recent advancements and open research challenges in the field, emphasizing the need for scalable, efficient, and provably secure aggregation methods to realize the full potential of EFL in sensitive domains like healthcare and finance.