acceptodds
Withdrawn from ICLR 2027

Forgetting without Seeing: Blind Federated Unlearning with Pure Random Noise

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated unlearning aims to remove the influence of a designated client from a trained global model while preserving model utility on the retained clients. Existing methods presuppose a fixed degree of unlearning client participation, whereas client availability is unpredictable in practice. Therefore, we formalize this setting as blind federated unlearning, where the unlearning client may never participate or may depart at any round, and the remaining rounds are completed solely by the retained clients. We find that when the unlearning client departs after partial participation, continuing with retained clients alone can substantially reverse the unlearning already achieved, revealing the need for persistent unlearning signals after client departure. To provide persistent surrogate signals without relying on semantic information about the unlearning data or auxiliary data, we propose Noise-as-Proxy Erasure (NoPE), which uses pure random noise generated at the server as proxy data. NoPE employs distance-based loss weighting to regulate the contribution of individual noise samples, enabling fine-grained control over the unlearning process. Extensive experiments demonstrate that NoPE achieves effective unlearning while maintaining competitive performance on retained data across different blind federated unlearning scenarios, and can be attached to existing federated unlearning methods to continue unlearning after client departure.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Learning, Federated Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wmcyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies blind federated unlearning, where the target client may be unavailable from the outset or depart before unlearning is complete. In the paper, the authors propose NoPE, which combines server-side training on class-conditioned random noise with retained-client updates.

Critical strengths
  1. The paper consider an understandable and practical research question, where the unlearning client can not participate during the unlearning phase.
  2. The method does not require client data for the unlearning phase.
Critical weaknesses
  1. The paper presentation require significant improvement.
  2. Although current related works have been investigated widely, parts of them are not yet peer-reviewed or does not from flagship venues like CVPR, NeurIPS, ICLR or ICML. Another speaking, it would be good if the literature reviews should be much more carefully considered. More specifically, more works in recent flagship venues should be investigated.
  3. One of my biggest concerns lay in the utilization of the mean and standard deviation of the input data for the augmentation. It is not well proven of how it can work. A comprehensive experimental evaluations and other visualization and analysis should be made to prove this claim. The paper can consider the literature reviews in this type of approach to indirectly improve the soundness of the paper.
  4. Beside the utilization of mean and standard deviation of the input data, NoPE combines training on a proxy task with recovery on retained data. This simplicity is somehow good, but the current evidence does not establish why the procedure removes the target client’s influence specifically. The noise proxy depends on the target client only through its dataset size and aggregation weight, so the server would inject the same signal whichever client requested unlearning.
  5. According to the experimental evaluations, the current manuscript should consider a more challenging dataset and more baselines in 2025 and 2026 to prove the method better.
Questions for authors
  1. What is the difference between prototype and mean of the data in the paper?
  2. As in the T-SNE figure, the clean and noisy data according to different class are not correlated. How do we achieve the generalized unlearning performance on the clean dataset when we apply it on the noisy dataset? It makes me feel confused about the technicality and soundness of the paper.
Reviewer_kcBxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a promissing and simple method with strong reported backdoor-removal results. However, the evidence does not yet establish its broader client-unlearning claims. And several experimental controls are missing.

Critical strengths
  1. Client departure during unlearning is practically relevant, and Figure 1 provides a clear motivation for examining what happens after departure.

  2. NoPE avoids a generator, auxiliary real data, and continued access to the departing client.

Critical weaknesses
  1. In the abstract, the paper claims that "continuing with retained clients alone can substantially reverse the unlearning already achieved." I don't think the claim is correct, for example, as their backdoor removal experiments, after unlearning a backdoor, and the client departs now, why can the unlearning be reversed? The remaining data do not have the backdoor data.

  2. In the introduction, the blind federated unlearning setting is insufficient. The assumption that the requesting client will depart before unlearning is completed is concerning. A client seeking removal of its data would ordinarily have an incentive to remain available until completion is confirmed.

  3. Most experiments measure suppression of a single backdoor implanted in all target-client samples. Low ASR does not establish removal of the client’s broader influence. Stronger clean-data evaluations and comparisons with retraining are needed.

  4. ASR rebound during clean retained-client training may indicate reactivation of a residual backdoor rather than reversal of successful information removal. The authors should distinguish behavioral suppression from actual removal and investigate the mechanism behind this rebound.

Questions for authors
  1. Table 5 is difficult to interpret: retraining retains approximately 75% attack accuracy, and several accuracy–recall pairs are incompatible with a balanced evaluation set. The authors should clarify sampling, class balance, thresholds, and attack adaptation, and report AUROC and TPR at fixed FPR.

  2. The paper motivates NoPE through rebound after unlearning signals stop, but does not apply the same test to NoPE. The authors should disable the noise proxy after completion and evaluate whether subsequent retained-client training restores backdoor behavior.

Reviewer_avTjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Noise-as-Proxy Erasure (NoPE), a federated unlearning method for settings where the client requesting removal is unavailable from the outset or departs during the unlearning process. The key observation is that continued training on retained clients can reactivate backdoor behavior that earlier unlearning rounds had suppressed. Based on this, NoPE replaces the unavailable client’s update with a server-side update trained on class-conditional Gaussian noise, using distance-based sample weights to preserve model utility. Experiments on CIFAR-10, CIFAR-100, and AGNews show that NoPE generally achieves a favorable trade-off between backdoor suppression and retained accuracy, both as a standalone method and as a continuation mechanism for existing federated unlearning approaches.

Critical strengths
  1. This paper is well motivated and presents a coherent progression from an empirical observation to the proposed method. In particular, the finding that retained-client training can reverse previously achieved backdoor suppression highlights a practical limitation of evaluating unlearning only at the moment the target client departs. This provides a useful motivation for maintaining an unlearning signal after client departure.
  2. The proposed method is simple and easy to integrate into existing federated learning pipelines. NoPE constructs its proxy data through random sampling without requiring semantic information about the target client’s data or an external auxiliary dataset. The proxy update uses standard FedAvg aggregation and can be attached to different unlearning methods without modifying their procedures before client departure.
  3. The empirical evaluation covers several relevant settings, including complete client absence, early departure, heterogeneous data partitions, varying retained-client participation, and both image and text classification. The main results show strong backdoor suppression while preserving substantially higher retained accuracy than several competing methods. The proxy-data ablations also provide an interesting observation that semantically related proxy data are not necessary to obtain low ASR in the evaluated settings.
Critical weaknesses
  1. The evidence for general client-level unlearning is less conclusive than the evidence for backdoor suppression. Most experiments introduce a fixed trigger into all training examples of the target client and evaluate whether the resulting behavior disappears. Removing this association does not necessarily establish that the remaining influence of the client’s data has been removed. Although the clean-data membership-inference experiment is a useful addition, the paper does not sufficiently establish equivalence to retraining. Moreover, given that unlearning reversal is the central motivation, it is unclear whether the achieved suppression persists after NoPE is stopped and training continues only on retained clients.
  2. The membership-inference evaluation requires clarification before its results can be interpreted confidently. In Table 5, NoPE has an IID MIA accuracy of 75.9% and recall of 30.9%. If these metrics are computed on the same balanced member/non-member test set, this combination is impossible: even with 100% specificity, accuracy would be at most 65.45%. An imbalanced evaluation set or different populations for the two metrics could explain these values, but the paper does not specify these details. The retraining MIA accuracy of 74.8%, despite the table indicating 50% as the desired reference, further complicates the interpretation.
  3. The distance-based weighting ablation does not clearly separate the effect of sample preference from the effect of changing optimization strength. Equation 3 divides the weighted loss by the number of samples rather than the sum of weights. Consequently, changing from uniform weights to distance-based weights changes both the relative contributions of samples and the overall loss scale. The improved retained accuracy may therefore partly reflect a weaker proxy update. The competitive performance of squared center-far weighting also suggests that additional controls are needed to substantiate the specific benefit of emphasizing center-near samples.
Questions for authors
  1. After the prescribed NoPE rounds are completed, what happens if the noise proxy is removed and training continues solely on retained clients? Could the authors report the subsequent ASR and retained accuracy, alongside a retraining reference under the same continuation protocol? Since the paper is motivated by unlearning reversal, this experiment would help establish whether NoPE achieves persistent removal or suppression that depends on continued proxy updates.
  2. Could the authors specify the member/non-member sample counts, sampling distributions, attack threshold, and evaluation populations used for MIA accuracy and recall in Table 5? In particular, how should the IID accuracy/recall pair of 75.9%/30.9% be reconciled, and why does the retrained model retain a 74.8% attack accuracy? Confusion matrices and a threshold-independent metric would help clarify whether the reported improvements reflect reduced membership information.
  3. Does the advantage of squared center-near weighting remain when the weighting variants have matched average weights or separately tuned proxy learning rates? A further useful control would randomly reassign the same weights to samples within each pseudo-class, preserving their overall scale while removing their relationship with distance. This would help determine whether the improvement specifically comes from selecting center-near samples.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.