A Comprehensive Investigation of Federated Unlearning: Challenges, Methods and Future Prospects in Privacy-Sensitive Applications

Wei Zhang

2024

Abstract

This paper reviews a range of federated unlearning techniques, with a focus on their applications, limitations, and potential benefits. Federated unlearning addresses privacy concerns by enabling the removal of specific data from machine learning models without requiring full retraining. This is particularly relevant in complying with legal regulations like General Data Protection Regulation (GDPR). Methods like FedEraser and FedCIO provide effective data removal by partitioning and clustering data, making them suitable for handling complex, non-independent and identically distributed (Non-IID) data. FedRecovery offers high precision by storing and rolling back model gradient updates, while other approximate methods such as F2UL optimize computational efficiency through differential privacy, striking a balance between privacy and performance. The analysis reveals the trade-offs between these exact and approximate methods, with the former ensuring better data removal precision but at a higher computational cost, and the latter being more resource-efficient but involving potential privacy risks. It can be concluded that future research should focus on developing standardized evaluation metrics, improving computational efficiency, and enhancing the adaptability of federated unlearning techniques to better manage Non-IID data in real-world applications. This research aims to guide advancements in federated unlearning, promoting its application in dynamically adaptive, privacy-sensitive machine learning scenarios.

Paper Citation

in Harvard Style

Zhang W. (2024). A Comprehensive Investigation of Federated Unlearning: Challenges, Methods and Future Prospects in Privacy-Sensitive Applications. In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-754-2, SciTePress, pages 569-573. DOI: 10.5220/0013528500004619

in Bibtex Style

@conference{daml24,
author={Wei Zhang},
title={A Comprehensive Investigation of Federated Unlearning: Challenges, Methods and Future Prospects in Privacy-Sensitive Applications},
booktitle={Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2024},
pages={569-573},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013528500004619},
isbn={978-989-758-754-2},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML
TI - A Comprehensive Investigation of Federated Unlearning: Challenges, Methods and Future Prospects in Privacy-Sensitive Applications
SN - 978-989-758-754-2
AU - Zhang W.
PY - 2024
SP - 569
EP - 573
DO - 10.5220/0013528500004619
PB - SciTePress