A Survey on Federated Unlearning: Challenges, Methods, and Future Directions

TL;DR

This survey formalizes federated unlearning (FU) as a principled extension of machine unlearning (MU) to federated learning (FL), addressing data-erasure requests under RTBF in decentralized settings. It introduces a unified FU workflow and a taxonomy based on who unlearns, what is unlearned, and how verification is performed, then surveys passive and active FU methods with server- and client-side verification. The paper analyzes FL-tailored optimizations, memory/communication/computation trade-offs, heterogeneity, and security threats, and discusses privacy-preserving approaches, verification, and future directions such as proofs of unlearning and integration with MLaaS. Together, these insights guide researchers and practitioners in building robust, verifiable FU systems that respect privacy regulations while maintaining FL performance across diverse domains.

Abstract

In recent years, the notion of ``the right to be forgotten" (RTBF) has become a crucial aspect of data privacy for digital trust and AI safety, requiring the provision of mechanisms that support the removal of personal data of individuals upon their requests. Consequently, machine unlearning (MU) has gained considerable attention which allows an ML model to selectively eliminate identifiable information. Evolving from MU, federated unlearning (FU) has emerged to confront the challenge of data erasure within federated learning (FL) settings, which empowers the FL model to unlearn an FL client or identifiable information pertaining to the client. Nevertheless, the distinctive attributes of federated learning introduce specific challenges for FU techniques. These challenges necessitate a tailored design when developing FU algorithms. While various concepts and numerous federated unlearning schemes exist in this field, the unified workflow and tailored design of FU are not yet well understood. Therefore, this comprehensive survey delves into the techniques and methodologies in FU providing an overview of fundamental concepts and principles, evaluating existing federated unlearning algorithms, and reviewing optimizations tailored to federated learning. Additionally, it discusses practical applications and assesses their limitations. Finally, it outlines promising directions for future research.

Figures (9)

• Figure 1: Machine unlearning. Naive retraining, discarding the trained model and starting training from scratch with remaining data after unlearned data removal, is computationally intensive. Conversely, machine unlearning, which resumes training from the trained model through an unlearning process, is much more cost-effective. The objective of MU is to ensure that the unlearned model achieves a performance level on par with that of the retrained model.
• Figure 2: Federated unlearning. In contrast to machine unlearning algorithms, which are typically executed by a single entity, FU systems involve multiple entities, including the unlearned client, remaining clients, and the central server, any of whom can act as the unlearner, responsible for executing the unlearning algorithm. Furthermore, the unlearning target may encompass either an entire client or specific partial data from a target client.
• Figure 3: Illustrative organization of the paper.
• Figure 4: An unified federated unlearning workflow. This workflow outlines the timeline for learning, unlearning, and verification. When the FU system receives an unlearning request, it can follow either the passive unlearning approach, where the target client exits the system immediately, or the "Active unlearning" approach, where the target client chooses to stay and participate in the unlearning process. Unlearning requests can be initiated by either the unlearned client or the server for various purposes. Furthermore, the unlearning and verification roles can be performed by the server, the target clients, the remaining clients, or a combination of both.
• Figure 5: Taxonomy of federated unlearning schemes.

Theorems & Definitions (5)

• Definition 1
• Definition 2
• Definition 3
• Definition 4
• Definition 5