A Privacy-Preserving Federated Learning Framework Based on Homomorphic Encryption

Abstract

he development of artificial intelligence has been significantly influenced by the growing availability of large and diverse datasets, which has enabled the rapid evolution of machine learning and deep learning techniques. Initially, centralized learning was adopted as the dominant approach, requiring data to be collected and stored in a single location for model training. While effective in performance, this method raises serious concerns regarding data privacy and security, particularly in domains such as healthcare. To overcome these limitations, federated learning has emerged as a distributed alternative that allows collab- orative model training while keeping data localized. In this setting, only model updates are exchanged between clients and the central server. However, despite not sharing raw data, these updates can still leak sensitive information through heuristic attacks, such as gradient inversion techniques that reconstruct original inputs. In response to this threat, this thesis presents a federated learning framework enhanced with homomorphic encryp- tion to secure the exchange of model updates. The proposed approach combines structural and sensitivity-based strategies to apply encryption where it is most needed, aiming to achieve a balance between data protection and learning efficiency. The framework was evaluated through a series of controlled experiments designed to assess its effectiveness in preserving privacy while maintaining model accuracy. The results demonstrate the practical potential of the framework as a secure solution for privacy-aware collaborative learning.