Abstract

The right to be forgotten, as stated in most data regulations, poses an underexplored challenge in federated learning (FL), leading to the development of federated unlearning (FU). However, current FU approaches often face trade-offs between efficiency, model performance, forgetting efficacy, and privacy preservation. In this paper, we delve into the paradigm of Federated Client Unlearning to guarantee a client the right to erase the contribution or the influence, introducing the first FU framework in medical imaging. In the unlearning process of a client, the proposed Model-Contrastive Unlearning marks a pioneering step towards feature-level unlearning, and Frequency-Guided Memory Preservation ensures smooth forgetting of local knowledge while maintaining the generalizability of the trained global model, thus avoiding performance compromises and guaranteeing rapid post-training. We evaluate our FCU framework on two public medical image datasets, including Intracranial hemorrhage diagnosis and skin lesion diagnosis, demonstrating our proposed framework outperforms other state-of-the-art FU frameworks working, with an expected speed-up of 10-15 times compared with retraining from scratch. The code and organized datasets will be made public.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1632_paper.pdf

SharedIt Link: https://rdcu.be/dV54j

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72117-5_23

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1632_supp.zip

Link to the Code Repository

https://github.com/dzp2095/FCU

Link to the Dataset(s)

https://www.kaggle.com/datasets/shonenkov/isic2018 https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection

BibTex

@InProceedings{Den_Enable_MICCAI2024,
        author = { Deng, Zhipeng and Luo, Luyang and Chen, Hao},
        title = { { Enable the Right to be Forgotten with Federated Client Unlearning in Medical Imaging } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15010},
        month = {October},
        page = {240 -- 250}
}

Reviews

Review #1

Please describe the contribution of the paper

The authors delve into the paradigm of Federated Client Unlearning and present the first FU framework in medical imaging to ensure the right of a target client to remove the contribution of their data from a trained global model efficiently.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1. The proposed Model-Contrastive Unlearning marks a pioneering step towards feature-level unlearning, and Frequency-Guided Memory Preservation ensures smooth forgetting of local knowledge while maintaining the generalizability of the trained global model, thus avoiding performance compromises and guaranteeing rapid post-training.
2. The code and the organized datasets are available.
3. The experiments are sufficient to evaluate the effectiveness of the proposed method.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1. In Introduction, the authors said that to make better use of the information contained in the teacher network, since [26], most knowledge distillation shifted from output distillation to feature distillation [13, 33], showing superior performance on various tasks. The pros and cons of output distillation to feature distillation depends on the specific situation.
2. In Figure 1, MCU and FDMP are conducted alternatively. So please clarify how to switch between the MCU and FDMP.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Do you have any additional comments regarding the paper’s reproducibility?

N/A
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
1. The organization of this paper is good.
2. The experiments are sufficient to evaluate the effectiveness of the proposed method.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

Weak Accept — could be accepted, dependent on rebuttal (4)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1. The organization of this paper is good.
2. The experiments are sufficient to evaluate the effectiveness of the proposed method.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #2

Please describe the contribution of the paper

This paper discussed the federated unlearning in the medical task domain. Two main techniques are proposed including MCU and FGMP. They evaluate the approach on two real-world tasks.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

1, The research topic is interesting and practical in the real-world application; 2, The proposed two techniques are convincing and solid. 3, They also take the computational efficiency into consideration.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

1, Federated unlearning is not a new topic. Could you please clarify the main difference with the existing federated unlearning works? Not just the medical application. 2, Presentation quality needs to be improved, such as Figure 2.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
Do you have any additional comments regarding the paper’s reproducibility?

N/A
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

Please see the weakness part.

Also, multiple times of experiment run are expected to report the avg and STD.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

Weak Accept — could be accepted, dependent on rebuttal (4)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This is an interesting and practical research topic. The authors utilize the acceptable techniques to address the challenges. The presentation is easy to follow. The experiment results support their statement.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #3

Please describe the contribution of the paper

This paper introduces Federated Client Unlearning (FCU) to address the challenge of the right to be forgotten in Federated Learning (FL). The authors propose a novel FU framework for medical imaging, featuring Model-Contrastive Unlearning for feature-level unlearning and Frequency-Guided Memory Preservation for smooth forgetting while maintaining model performance. FCU outperforms existing FU frameworks on two medical image datasets, offering a speed-up of 10-15 times compared to retraining from scratch. Code and datasets will be made publicly available.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- This work offers a novel FU framework for medical tasks to protect patients’ data forgotten rights.
- Extensive case studies.
- Good writing.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- The privacy performance of the proposed framework needs further verification.
- In terms of the number and categories of unlearn samples, this paper needs to further explore the performance impact of these two parameters on the proposed method.
- Regarding the numerical results in Table 2, I still have some concerns.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Do you have any additional comments regarding the paper’s reproducibility?

NA
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
Comments:
1. In MU, membership inference attack is usually adopted to measure whether the proposed method has good privacy performance. However, in this paper, the authors seem to be missing an evaluation of the privacy performance of the method, which is essential for medical tasks. It would be better if the authors could provide theoretical or empirical analysis of privacy.
2. Secondly, the number of unlearning samples and the corresponding category labels are key parameters that affect unlearning performance. However, the authors seem to be missing an evaluation of these two parameters. I recommend that authors refer to the following references to expand the unlearning scenarios (e.g., unlearn labels and unlearn features) of this paper. [R1] Chen M, Zhang Z, Wang T, et al. When machine unlearning jeopardizes privacy[C]//Proceedings of the 2021 ACM SIGSAC conference on computer and communications security. 2021: 896-911. [R2] Warnecke A, Pirch L, Wressnegger C, et al. Machine unlearning of features and labels[J]. In Proc. of NDSS, 2021. [R3] Wang J, Guo S, Xie X, et al. Federated unlearning via class-discriminative pruning[C]//Proceedings of the ACM Web Conference 2022. 2022: 622-632.
3. The observed phenomenon of Retrain outperforming Origin in Tables 1 and 2 warrants further explanation to clarify the underlying reasons behind this unexpected outcome. Additionally, exploring the impact of the number of unlearned samples on unlearning performance is essential, as it directly influences the performance of both Retrain and unlearn models. Analyzing this aspect will provide valuable insights into how variations in the number of unlearned samples affect the overall performance of the proposed method and shed light on the underlying mechanisms driving these performance differences.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

Weak Accept — could be accepted, dependent on rebuttal (4)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This article does design an efficient FU framework for medical tasks, which is commendable. However, there are still some concerns that need to be addressed in terms of privacy performance and parameter analysis.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Author Feedback

We thank the reviewers for their comments and constructive suggestions on our manuscript. We are committed to revising our manuscript according to the suggestions, which we believe will significantly improve the quality and clarity of our work. Below, we explain our responses in detail.

Output distillation and feature distillation [R1] We appreciate your suggestion to clarify the pros and cons of output distillation and feature distillation. In this study, we proposed the first feature distillation-based unlearning method – Model-Contrast Unlearning (MCU). In our ablation study in section 3.2, we replaced our MCU with two output distillation-based methods [5][34]. Our results demonstrate that MCU can better preserve performance on the test set while achieving a similar forgotten error, as shown in Fig 2. In future studies, we will further explore and compare output distillation and feature distillation in unlearning to highlight the pros and cons of both strategies.

Clarification the way to switch between the MCU and FGMP [R1] We apologize for causing misunderstanding and promise to revise it in the final version. The revised description in section 2.3 is shown below for your convenience (revisions are indicated by [brackets]):
Specifically, we conduct [FGMP] every T_FGMP iterations (e.g., T_FGMP is set to 10 in our experiment) while [MCU is continuously executed], resulting in an unlearned model M_un.

Difference with the existing federated unlearning studies [R3] In this paper, we proposed Model-Contrastive Unlearning, which marks a pioneering step towards feature-level unlearning. To the best of our knowledge, no method has yet considered encouraging the student model to learn from the “Bad Teacher” at the feature level to guarantee a higher level forgetting, marking an opportunity for innovation in unlearning. This approach differs from other studies that focus on output distillation and shows better performance in our ablation study in section 3.2. Additionally, we introduced a novel Frequency-Guided Memory Preservation method that ensures smooth forgetting of local knowledge while maintaining the generalizability of the trained global model, thereby avoiding performance compromises and guaranteeing rapid post-training.
In the second paragraph of section 1 of our original manuscript, we discussed the main weaknesses of existing federated unlearning (FU) studies, highlighting their compromise of performance or privacy. Our method outperforms other state-of-the-art FU frameworks in terms of model performance and unlearning speed without compromising privacy.

Presentation quality of Fig. 2. [R3] We promise to improve it in our final version.

Further veriﬁcation of privacy performance [R4] We appreciate the advice on further verification methods, such as Membership Inference Attack (MIA) or empirical analysis of privacy performance. We will explore more evaluations of privacy performance and conduct empirical analysis in our future work.

Number and categories of unlearn samples [R4] In the FU review paper [25], the authors summarized three objectives in FU: Sample Unlearning, Class Unlearning, and Client Unlearning. Previous researchers usually focus on one of these objectives, as in the paper you mentioned. In our study, we focus on Client Unlearning, which is peculiar to federated settings where the client that requests the unlearning would like to erase the contribution of its entire local dataset.

Retrain outperform Origin in Table 2 [R4] To clarify, Retrain only outperforms Origin on one metric, i.e., accuracy, in Table 2. In our study, we have simulated heterogeneous multi-source data (non-IID setting) as described in Datasets in section 3.1. We appreciate the reviewer for raising this point. We will explore how the distribution and quality of unlearned data affect model performance in federated learning in future research.

Meta-Review

Meta-review not available, early accepted paper.

back to top

Enable the Right to be Forgotten with Federated Client Unlearning in Medical Imaging

Author(s):