CoChat
Sign Up
Cited by 202Privacy-Preserving Technologies in DataCloud Data Security SolutionsCryptography and Data Security

The Right to be Forgotten in Federated Learning: An Efficient Realization with Rapid Retraining

Open access
Liu, Yi, Xu, Lei, Yuan, Xingliang, Wang, Cong, Li, Bo·2022·IEEE INFOCOM 2022 - IEEE Conference on Computer Communications
DOI 10.1109/infocom48880.2022.9796721
Open this paper in CoChat — free accountOpen PDF

Abstract

In Machine Learning, the emergence of the right to be forgotten gave birth to a paradigm named machine unlearning, which enables data holders to proactively erase their data from a trained model. Existing machine unlearning techniques focus on centralized training, where access to all holders’ training data is a must for the server to conduct the unlearning process. It remains largely underexplored about how to achieve unlearning when full access to all training data becomes unavailable. One noteworthy example is Federated Learning (FL), where each participating data holder trains locally, without sharing their training data to the central server. In this paper, we investigate the problem of machine unlearning in FL systems. We start with a formal definition of the unlearning problem in FL and propose a rapid retraining approach to fully erase data samples from a trained FL model. The resulting design allows data holders to jointly conduct the unlearning process efficiently while keeping their training data locally. Our formal convergence and complexity analysis demonstrate that our design can preserve model utility with high efficiency. Extensive evaluations on four real-world datasets illustrate the effectiveness and performance of our proposed realization.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1109/infocom48880.2022.9796721

Cites

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The solution by Liu et al. [93] will result in a significant reduction in accuracy when dealing with complex models.

    “However, this solution will result in a significant reduction in accuracy when dealing with complex models.”

    From Machine Unlearning: A Survey (2024) · cites this paper as [93]

    The Evaluation (p. 9) reports that Liu et al.’s algorithm achieves accuracy comparable to retraining from scratch and says momentum compensates for the model-utility loss from its approximation. The Introduction (p. 2) describes the solution as model agnostic and reports only a small accuracy loss on CelebA, contrary to the claimed significant reduction for complex models. (LLM, conf=0.87)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • The second metric by Liu et al. [93] is the difference between the distribution of the model after the unlearning process and the distribution of the retraining model.

    “The second metric is the difference between the distribution of the model after the unlearning process and the distribution of the retraining model.”

    From Machine Unlearning: A Survey (2024) · cites this paper as [93]

    The Experiment Setup (p. 8) explicitly identifies the second evaluation metric as Symmetric Absolute Percentage Error between the test accuracies of the unlearned model and the retraining baseline, not a difference between their model distributions. The paper separately discusses approximately identical model distributions as an unlearning criterion (Formalizing the Problem, p. 4), but does not define that as its second metric. (LLM, conf=0.87)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • Liu et al. [93] transferred the unlearning method from a centralized environment to federated learning by proposing a distributed Newton-type model updating algorithm to approximate the loss function trained by the local optimizer on the remaining dataset.

    “Liu et al. [93] transferred the unlearning method from a centralized environment to federated learning by proposing a distributed Newton-type model updating algorithm to approximate the loss function trained by the local optimizer on the remaining dataset.”

    From Machine Unlearning: A Survey (2024) · cites this paper as [93]

  • Liu et al. [93] use diagonal empirical Fisher Information Matrix (FIM) to efficiently and accurately approximate the inverse Hessian vector, rather than computing it directly, to further reduce the cost of the retraining process.

    “They also use diagonal empirical Fisher Information Matrix (FIM) to efficiently and accurately approximate the inverse Hessian vector, rather than computing it directly, to further reduce the cost of the retraining process.”

    From Machine Unlearning: A Survey (2024) · cites this paper as [93]

  • The method by Liu et al. [93] is based on the Quasi-Newton method and uses a first-order Taylor expansion.

    “This method is based on the Quasi-Newton method and uses a first-order Taylor expansion.”

    From Machine Unlearning: A Survey (2024) · cites this paper as [93]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

The citation lists haven’t been fetched yet

References 40 · Cited by 202

Research works printed but not linked: 17

  1. [1][1] L. Bourtoule, V. Chandrasekaran, C. A. Choquette-Choo, H. Jia, A. Travers, B. Zhang, D. Lie, and N. Papernot, 'Machine unlearning,' in Proc. of IEEE S&P , 2019. -
  2. [4][4] P. W. Koh and P. Liang, 'Understanding black-box predictions via influence functions,' in Proc. of ICML , 2017. -
  3. [7][7] C. Guo, T. Goldstein, A. Hannun, and L. Van Der Maaten, 'Certified data removal from machine learning models,' in Proc. of ICLR , 2020. -
  4. [8][8] Z. Izzo, M. A. Smart, K. Chaudhuri, and J. Zou, 'Approximate data deletion from machine learning models,' in Proc. of AISTATS , 2021. -
  5. [9][9] M. Chen, Z. Zhang, T. Wang, M. Backes, M. Humbert, and Y. Zhang, 'Graph unlearning,' in Proc. of CCS , 2021. -
  6. [10][10] Y. Wu, E. Dobriban, and S. Davidson, 'Deltagrad: Rapid retraining of machine learning models,' in Proc. of ICML , 2020. -
  7. [11][11] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, 'Communication-efficient learning of deep networks from decentralized data,' in Proc. of AISTATS , 2017. -
  8. [22][22] J. Brophy and D. Lowd, 'Machine unlearning for random forests,' in Proc. of ICML , 2021. -
  9. [27][27] Y. Huang, S. Gupta, Z. Song, K. Li, and S. Arora, 'Evaluating gradient inversion attacks and defenses in federated learning,' in Proc. of NeurIPS , 2021. -
  10. [30][30] A. S. Berahas, J. Nocedal, and M. Tak´ aˇ c, 'A multi-batch l-bfgs method for machine learning,' in Proc. of NeurIPS , 2016. -
  11. [31][31] R. Bollapragada, J. Nocedal, D. Mudigere, H.-J. Shi, and P. T. P. Tang, 'A progressive batching l-bfgs method for machine learning,' in Proc. of ICML , 2018. -
  12. [33][33] J. Friedman, T. Hastie, R. Tibshirani et al. , The elements of statistical learning . Springer series in statistics New York, 2001, vol. 1, no. 10. -
  13. [35][35] S. Becker, Y. Le Cun et al. , 'Improving the convergence of backpropagation learning with second order methods,' in Proceedings of the 1988 connectionist models summer school , 1988, pp. 29-37. -
  14. [36][36] S. Zheng, Q. Meng, T. Wang, W. Chen, N. Yu, Z.-M. Ma, and T.-Y. Liu, 'Asynchronous stochastic gradient descent with delay compensation,' in Proc. of ICML , 2017. -
  15. [37][37] D. P. Kingma and J. Ba, 'Adam: A method for stochastic optimization,' in Proc. of ICLR , 2015. -
  16. [38][38] T. Tieleman, G. Hinton et al. , 'Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude,' COURSERA: Neural networks for machine learning , vol. 4, no. 2, pp. 26-31, 2012. -
  17. [41][41] Z. Yao, A. Gholami, Q. Lei, K. Keutzer, and M. W. Mahoney, 'Hessianbased analysis of large batch training and robustness to adversaries,' in Proc. of NeurIPS , 2018. -

Authors
Liu, Y., Xu, L., Yuan, X., Wang, C., & Li, B.
Affiliations
  • City University of Hong Kong (HK)
  • Nanjing University of Science and Technology (CN)
  • Monash University (AU)
  • Hong Kong University of Science and Technology (HK)
ORCID
Venue
IEEE INFOCOM 2022 - IEEE Conference on Computer Communications
Publisher
IEEE
Published
2022-05-02
Pages
1749-1758
DOI
10.1109/infocom48880.2022.9796721
OpenAlex
W4221165231
Cited by
202
References
40
Citations by year
2026: 38 · 2025: 74 · 2024: 62 · 2023: 25 · 2022: 3
FWCI
22.65
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE Conference on Computer Communications (INFOCOM) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Natural Science Foundation of China
  • Natural Science Foundation of Jiangsu Province
  • Australian Research Council
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Information Systems › Cloud Data Security Solutions
  • Physical Sciences › Computer Science › Artificial Intelligence › Cryptography and Data Security
Keywords
right to be forgotten, machine unlearning, federated learning, model utility preservation, convergence analysis, computational complexity

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review