Federated Unlearning

By Abhivardhan

Definition 

Federated Unlearning is a technique that enables a machine learning model trained through Federated Learning to forget the influence of a particular client’s data. Aligned with the right to be forgotten in the digital space, it provides a mechanism through which contributors or clients in a federated learning system can seek removal of their data’s influence from the global model. This technique is also useful where the underlying data becomes outdated, invalid, or otherwise detrimental to the accuracy and responsiveness of the Federated Learning model.

Commentary

Origin of the term 

The concept of Federated Unlearning was first introduced by Gaoyang Liu, Yang Yang, Xiaoqiang Ma, Chen Wang, and Jiangchuan Liu in their 2020 paper, ‘Federated Unlearning’.¹ Building upon the concept of Federated Learning, the authors addressed the challenge of enabling a federated model to "forget" the contribution of a client whose data is sought to be removed, without requiring the entire model to be retrained from scratch. To achieve this, they proposed FedEraser,² the first practical federated unlearning algorithm, which reconstructs the global model by calibrating and aggregating the historical model updates of the remaining clients, thereby removing the influence of the unlearned client's data in a computationally efficient manner.

Operation in Practice

For a practical example, the Fraunhofer Institute for Software and Systems Engineering ISST, in collaboration with Fujitsu Research, has developed a federated unlearning solution for decentralised AI collaborations involving corporate data. The method allows a company leaving a collaborative AI project to remove the influence of its proprietary data from the shared model by tracing the model’s training history to the point where that company’s data was introduced and resuming training from there without such data. This enables a cleaner removal of corporate data while being more efficient than retraining the entire model from scratch.³

Sector 

Federated Unlearning finds application across sectors such as healthcare, defence, financial services, insurance, marketing, and mobile applications, among others.⁴

Relevance vis-a-vis Competition Law 

While primarily a privacy-preserving mechanism, Federated Unlearning has indirect relevance to competition law, as it ensures that the clients can have their corporate data removed from the model upon exit.⁵ Such assurance alleviates any perpetual free-riding concerns among the data-rich enterprises that may otherwise hesitate in entering federated learning arrangements, eventually leading to more consumer and firm welfare.

Alternative Nomenclature

Federated Unlearning is also referred to as Federated Machine Unlearning,⁶ or as the Right to be Forgotten in Federated Learning. ⁷

1. ---

   1. Gaoyang Liu, Yang Yang, Xiaoqiang Ma, Chen Wang and Jiangchuan Liu, 'Federated Unlearning' (arXiv, 27 December 2020) <https://arxiv.org/abs/2012.13891> accessed 13 August 2026.↩︎

   2. Gaoyang Liu, Xiaoqiang Ma, Yang Yang, Chen Wang and Jiangchuan Liu, 'FedEraser: Enabling Efficient Client-Level Data Removal from Federated Learning Models' (2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), IEEE 2021) <https://ieeexplore.ieee.org/document/9521274> accessed 13 August 2026.↩︎

   3. Fraunhofer-Gesellschaft, 'Removing Corporate Data from AI Models' (idw – Informationsdienst Wissenschaft, 2 April 2026) <https://nachrichten.idw-online.de/2026/04/02/removing-corporate-data-from-ai-models> accessed 13 August 2026.↩︎

   4. Duality Technologies, 'Federated Learning Applications' (Duality Technologies Blog) <https://dualitytech.com/blog/federated-learning-applications/ > accessed 13 August 2026.↩︎

   5. Jiajun Meng, Jing Chen, Dongfang Zhao and Lin Liu, 'Federated Learning and Free-riding in a Competitive Market' (arXiv, 16 October 2024) <https://arxiv.org/pdf/2410.12723> accessed 13 August 2026↩︎

   6. Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou and Philip S Yu, 'Update Selective Parameters: Federated Machine Unlearning Based on Model Explanation' (arXiv, 18 June 2024) <https://arxiv.org/abs/2406.12516> accessed 13 August 2026.↩︎

   7. Yi Liu, Lei Xu, Xingliang Yuan, Cong Wang and Bo Li, 'The Right to be Forgotten in Federated Learning: An Efficient Realization with Rapid Retraining' (arXiv, 14 March 2022) <https://arxiv.org/abs/2203.07320> accessed 13 August 2026.↩︎