Enhancing Privacy and Utility in Federated Learning: A Hybrid P2P and Server-Based Approach with Differential Privacy Protection
TL;DR #
Enhancing Privacy and Utility in Federated Learning: A Hybrid P2P and Server-Based Approach with Differential Privacy Protection explores a federated architecture that combines peer-to-peer communication with server-based aggregation.
Problem #
The central server in a traditional federated-learning system simplifies coordination but can become a privacy and scalability bottleneck. Differential privacy can protect updates, but the additional noise may reduce model utility.
What did we do? #
We designed a hybrid architecture in which clients can exchange information through peer-to-peer connections while retaining server-based coordination where it is useful. We also studied how differential-privacy protection affects model performance in this setting.
The experiments compare privacy and utility under different configurations and analyze how the architecture changes the cost of privacy protection.
Main findings #
Federated learning does not have to rely on a single communication pattern. Combining peer-to-peer and server-based mechanisms can offer another point in the design space for balancing privacy, communication, and model utility.
Resources #
Cite this paper #
@conference{corbucci2024hybrid,
author={Luca Corbucci and Anna Monreale and Roberto Pellungrini},
title={Enhancing Privacy and Utility in Federated Learning: A Hybrid P2P and Server-Based Approach with Differential Privacy Protection},
booktitle={Proceedings of the 21st International Conference on Security and Cryptography - Volume 1: SECRYPT},
year={2024},
pages={592-602},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012863600003767},
isbn={978-989-758-709-2}}
}