Δ SFL: (Decoupled Server Federated Learning) to Utilize DLG Attacks in Federated Learning by Decoupling the Server

Sudipta Paul, Vicenç Torra

2023

Abstract

Federated Learning or FL is the orchestration of centrally connected devices where a pre-trained machine learning model is sent to the devices and the devices train the machine learning model with their own data, individually. Though the data is not being stored in a central database the framework is still prone to data leakage or privacy breach. There are several different privacy attacks on FL such as, membership inference attack, gradient inversion attack, data poisoning attack, backdoor attack, deep learning from gradients attack (DLG). So far different technologies such as differential privacy, secure multi party computation, homomorphic encryption, k-anonymity etc. have been used to tackle the privacy breach. Nevertheless, there is very little exploration on the privacy by design approach and the analysis of the underlying network structure of the seemingly unrelated FL network. Here we are proposing the ΔDSFL framework, where the server is being decoupled into server and an analyst. Also, in the learning process, ΔDSFL will learn the spatio information from the community detection, and then from DLG attack. Using the knowledge from both the algorithms, ΔDSFL will improve itself. We experimented on three different datasets (geolife trajectory, cora, citeseer) with satisfactory results.

Paper Citation

in Harvard Style

Paul S. and Torra V. (2023). Δ SFL: (Decoupled Server Federated Learning) to Utilize DLG Attacks in Federated Learning by Decoupling the Server. In Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT; ISBN 978-989-758-666-8, SciTePress, pages 577-584. DOI: 10.5220/0012150700003555

in Bibtex Style

@conference{secrypt23,
author={Sudipta Paul and Vicenç Torra},
title={Δ SFL: (Decoupled Server Federated Learning) to Utilize DLG Attacks in Federated Learning by Decoupling the Server},
booktitle={Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT},
year={2023},
pages={577-584},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012150700003555},
isbn={978-989-758-666-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT
TI - Δ SFL: (Decoupled Server Federated Learning) to Utilize DLG Attacks in Federated Learning by Decoupling the Server
SN - 978-989-758-666-8
AU - Paul S.
AU - Torra V.
PY - 2023
SP - 577
EP - 584
DO - 10.5220/0012150700003555
PB - SciTePress