acceptodds
Withdrawn from ICLR 2027

DeRe: Controlling Update Reintroduction in Cross-Domain Federated Unlearning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated unlearning (FU) aims to remove requested training effects from a federated model without retraining from scratch. Some FU requests go beyond excluding data from training and require the resulting global model to have low accuracy on the requested data while preserving retained knowledge. We extend this goal to cross domain FU, a more practical setting where clients may have different input distributions. After unlearning, the global model should suppress the requested classes on the forget client's domain while preserving the same classes on other domains and performance on the remaining data. In this setting, class information from other clients and domain information retained by the forget client may combine during aggregation and restore the target behavior. We refer to this recovery path as Update Reintroduction (UR). Stronger deletion can reduce UR but may also hurt retained accuracy. To address this tension, we propose DeRe (De-Reintroduction) with three stages. Probe Based Mask Selection identifies the edit region using relative deletion responses. Projected Class Level Unlearning coordinates deletion and retained updates at the client and server. Mask Constrained Recovery restores utility outside the selected mask without overwriting the selected edits. Experiments on Digit5, DomainNet, and PACS show that DeRe improves the balance between target suppression and retained accuracy for both single and multi class requests. Code is available at https://anonymous.4open.science/r/wqeimhaikc.

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

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_dZLYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a novel method for federated domain-specific class unlearning, called DeRe. The paper appears to address a non-standard unlearning setting: it assumes that clients belong to different domains while sharing the same set of classification classes. When a class-unlearning request is issued by a client, the global model should be updated so that it no longer assigns samples from that client’s domain to the requested class, while samples of the same class from other clients’ domains should still be correctly classified. Differently from a “traditional” retrained model, in the proposed setting, a forced misclassification for all the samples to forget is expected.

The paper proposes a three-stages mechanism and needs a recovery phase. The authors assume that the clients and server are trusted; part of the unlearning phase requires the requesting client to actively participate.

The proposed method is evaluated against six state-of-the-art baselines which, however, are not designed to solve the same problem described in the paper.

Critical strengths
  1. The paper defines and addresses what seems a novel federated domain-specific class unlearning setting.
Critical weaknesses
  1. In experiments, authors compare to baselines that are designed to approximate the performance or behavior of a "retrained model"; however, if such "traditional" retrained model is not the gold standard baseline, most of the results may appear inflated by optimizing for a different objective (state-of-the-art baselines try to match retrained-model performances which do not represent the best possible within the unlearning setting of the paper, while DeRe is built to optimize what is expected from assumptions). As a result, most of the baselines in main tables (Table 1 and Table 2) report "bad" performances for the ΔRF\Delta_{RF} metric used as principal axis.

  2. Deliberately forcing the model to fail on domain-specific samples may create a very obvious behavioral signature that could itself reveal information about the deletion request or about the forgotten domain. Such forced over-forgetting can become a membership/unlearning signal.

Questions for authors

Clarification requests:

  1. The paper, in light of its non-standard unlearning setting, introduces a new metric, ΔRF\Delta_{RF}; authors state that they follow standard practice of previous work, but I did not find evidence of this metric being used. Authors then proceed to evaluate the best performing method based on this metric. Can the authors clarify what is standard within existing FU literature and what is newly introduced in their paper?

  2. Deliberately forcing the model to fail on domain-specific samples may create a very obvious behavioral signature that could itself reveal information about the deletion request or about the forgotten domain. Such forced over-forgetting can become a membership/unlearning signal. Can the authors discuss on this?

  3. Can the authors provide an analysis of the forced missclassification: is this misclassification inflating one specific class or a specific subset of classes, or it redistributes the classification across all the others?

  4. From Figure 5, the method seems insensitive to its hyperparameter; I think that this analysis should cover ranges of tuning where it is detrimental for performances. Can the authors show a larger landscape of hyperparameters including tuning that makes the method underperform?

  5. No space is dedicated to the assumption of honest clients and server (defined as "trusted server and clients"): the clients have to actively participate to the unlearning process (and have to run a specific local algorithm). This assumption may be defensible, but limits the scope of the proposed solution (e.g., can it be used in regular cross-device FL settings?). Can the authors detail on this or explicitly acknowledge what limitations such assumptions imply?

  6. Can the authors report variance for their experiments?

  7. Can the authors report the complete detail for reproduction of other baselines results? How have they been tuned?

  8. In Section 4.4, the paper defines θpost\theta_{post} as the unlearned model after TrecT_{rec} “recovery rounds.” However, the recovery protocol used to obtain the results in Tables 1 and 2 remains unclear from Section 5.2. In particular: (a) is TrecT_{rec} fixed and identical across all baselines, or is a common total communication/computation budget enforced, considering that different methods have different unlearning costs? (b) What criterion is used to stop the recovery phase?

Suggestions for future work:

  • Indicate the model architectures used in experiments in the main text and not only in appendix.
Reviewer_NV78OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Federated unlearning removes requested training effects from a trained model. This paper considers requests that name a class: given the pair (f,c)(f, c), the global model should suppress class c on the forget client's domain while class c remains recognizable on the other clients' domains.

The paper names a recovery path, Update Reintroduction (UR), in which class information from other clients recombines with domain information retained by f. DeRe addresses this recovery path in three stages: a probe-scored parameter mask (PBMS), projected coordination of deletion against retention (PCLU), and recovery restricted to outside the mask (MCR). Experiments cover Digit5, DomainNet and PACS with Retrain as the reference and seven FU baselines.

Critical strengths

S1. Class-level deletion where the class stays on the other clients. The request removes one class from one client while every other client keeps that class, so the retained data still contains it and the request is hard rather than trivial. It is specific, it maps onto a plausible deployment case such as a hospital withdrawing one diagnostic class after a policy change, and it makes a falsifiable prediction: that suppressing the class and preserving it elsewhere pull against each other. The paper also treats retraining as a data-deletion reference rather than a behavioural oracle, since a model retrained without the requested samples still recognises the class on the forget client's domain through cross-domain transfer.

S2. Joint reporting of forgetting and retained accuracy. Each table reports the accuracy on the requested class together with the accuracy on everything retained, and the derived quantities summarise the gap between them, so the trade-off is the unit of comparison rather than one side of it. In the single-class results two methods reach zero on the requested class on all three datasets with retained accuracy of 86.30, 60.54 and 63.32 against 86.95, 61.92 and 68.47. Since erasure alone does not separate them, the retained-accuracy column is what does.

S3. Two request sizes, with the baselines run under one protocol. A single-class request and a multi-class one are evaluated, the latter averaged over the requested classes within one unlearning procedure rather than run as several sequential requests. All methods start from the same pretrained model, and the baselines run from their released code under the same client splits, request and privacy rule. Membership-inference success and time-to-forget are reported beside accuracy. The ablations are cumulative stage additions, and the paper reports that the projected update coordination (PCLU) provides the main utility gain, with the recovery stage reducing the loss that remains.

Critical weaknesses

W1. The evaluation does not test the paper's core points.

The cross-domain half of the goal is asserted rather than scored: the claim that the class stays recognizable on other domains has no separate number, and every experiment uses the same client and the same class, so the request stays fixed. Update Reintroduction is likewise inferred rather than measured, its only support being a retraining diagnostic whose own text says "consistent with", with no tracking of whether reintroduction appears during unlearning and recovery. Table 5 is misread on the same point: dropping the update coordination costs 0.56 points of retained accuracy on Digit5 but 42.78 and 35.59 on DomainNet and PACS, which points to coordination carrying the utility, while the paper reads the opposite conclusion out of the same table.

The headline number cannot carry the comparison built on it. Section 5.3 sets the method against FedAU on the shared zero and then claims the win on retained accuracy, which is the dimension where the two actually differ; reaching zero is not an advantage when the method compared against reaches it too. The related work does not engage the published class-level methods or explain how this request differs from them. FUCRT (ICCV 2025) reports erasure of the unlearning classes with preserved accuracy on the remaining ones under non-IID partitions, so a comparable evaluation is already in print and the literature review does not reach it. The authors should either bring that line of work into the comparison or state how their setting and result differ from it; leaving both out weakens the novelty claim rather than protecting it.

W2. The ablation measures the wrong component against the claim.

The abstract states that the ablations support PBMS "for target suppression", and Section 5.4 concludes that response-based selection "is needed to locate deletion specific groups". Table 5 does not support either reading: the forgetting accuracy is identical at 98.78, 64.18 and 55.26 across the base row, the base row with PBMS added, the same with the update coordination added, and the full method, so within the cumulative series PBMS changes target suppression by nothing at all. What PBMS does change is the retained-accuracy loss, by 7.85 points on DomainNet and 4.61 on PACS and by 0.06 on Digit5. The component that affects suppression is therefore being credited for an effect the table attributes to a different axis.

A row that isolates PBMS is also missing, and the reference point for the whole table is undefined. "Base" appears once in the paper, as a row label, with nothing anywhere saying what it contains. The single row where suppression drops is the one that replaces PBMS with a random mask, and that substitution changes which parameters are edited rather than testing whether PBMS's selection rule contributes anything on its own. Defining the base and adding a variant with the mask omitted would settle what PBMS is worth.

The same problem of scale applies to the comparisons the paper does make. The margins it claims are 0.22 points of retained accuracy over Retrain on DomainNet and 0.14 over Retrain on Digit5 in the multi-class setting, drawn from a single request per dataset with three seeds reported as a mean, no dispersion anywhere, hyperparameters that change per dataset, and no multi-class configuration stated. Margins of that size cannot be read as a method effect while none of those quantities is reported.

W3. The objective is optimised past the retrain floor without evidence that the trade pays.

The paper states that a model retrained on D_ret still recognises c on f's domain and then drives the accuracy on the requested class to zero anyway. That goes beyond what data deletion alone leaves, and the methods reaching zero are separated only by what they give up: 4.50 points of retained accuracy against Retrain on Digit5, 1.72 on PACS, and 0.22 gained on DomainNet. Membership inference does not support the trade either, since the signal is higher than Retrain's on DomainNet (0.517 against 0.474) and PACS (0.564 against 0.491). If suppressing the class below the level a retrained model reaches is the goal, the paper needs to say what that buys, and its own metrics currently argue the other way.

Questions for authors
  1. The stated goal needs the class to stay recognizable on the other clients' domains, but that half is scored only through an aggregate over all retained data. Can class-level accuracy be reported on each non-forget client's domain, and on the forget client's remaining classes, as separate figures?

  2. Update Reintroduction is the mechanism the method is named after, and its support is a retraining diagnostic. What happens to it when one of its two ingredients is removed, for instance when class c is withheld from the non-forget clients, or when the forget client's retained classes are withheld?

  3. Removing the update coordination costs 0.56 points of retained accuracy on Digit5 but 42.78 and 35.59 on DomainNet and PACS. If reintroduction is the explanation, why is its cost so uneven across datasets, and is that asymmetry being used?

  4. What is the Base row, and on which axis is PBMS expected to matter? The forgetting accuracy is unchanged across the cumulative rows while the retained-accuracy loss moves, so a variant with the mask omitted would show whether the selection rule contributes to suppression or to utility.

  5. Given that a model retrained on the retained data already recognises the class on the forget client's domain, what does suppression below that level buy, and which of the claims in the paper would the authors keep if it buys nothing measurable on either retained accuracy or membership inference?

Reviewer_atQpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies class unlearning in cross-domain FL. The forget client wants class c suppressed on its own domain, while other clients still have c and the model should keep recognizing it there. The authors show that even retraining without the deleted data still recognizes c on the forget domain, and they call this "Update Reintroduction". DeRe tackles it with a probe-based neuron mask, masked deletion with gradient projection, and a recovery phase that only updates parameters outside the mask.

Critical strengths
  • Fig. 2 is a nice observation. Retraining without the data still gets about 42% on DomainNet and 25% on PACS for the deleted class, so deleting data and suppressing the class are clearly not the same thing.
  • The trust model is stated clearly, and all methods start from the same pretrained model.
  • The random mask ablation with the same per-layer budget is a good control.
Critical weaknesses
  1. I'm not convinced this is unlearning. Retrain is the deletion reference, but DeRe is rewarded for going way below it (0% vs 25-98%). The hospital privacy example only needs the data removed, and retrain already does that. Also, a model at 0% where retrain is at 25-98% is easy to tell apart from retrain, so the request itself becomes visible.

  2. The main claim isn't actually measured. RA mixes all classes and domains, so it can hide damage to class c on the other domains. I'd want per-domain class c accuracy and accuracy on the forget client's other classes. UR is also never tested after unlearning. Nobody continues normal FL training after DeRe to check whether the class comes back, and recovery is only 1-3 rounds.

  3. The comparison seems tilted. Most baselines (Retrain, FCU, FOUL, FUSED) are built for data or client removal, so the RA - FA metric punishes them by design. FOUL is a client unlearning method and I couldn't find how it was adapted. FedOSD at 8% RA on Digit5 and FOUL at 33% RA on PACS look like broken runs. DeRe is tuned per dataset (Table 7), and I see no sign the baselines got the same treatment.

  4. The gains over FedAU are small. Both reach 0 FA, the RA gap is 0.65 and 1.4 points on Digit5 and DomainNet, and FedAU is about 40x faster. Everything is from one request (client 0, class 0) with no std. The PACS forget set looks like about 76 images, since every FA value in Table 1 is a multiple of 1/76. That's odd for a mean over 3 seeds. In the multi-class case DeRe doesn't reach 0, and it loses more RA than Retrain on DomainNet.

  5. The novelty is limited and some close work is missing. PBMS is basically a forget/retain importance ratio, like SSD (Foster et al., AAAI 2024) and SalUn (Fan et al., ICLR 2024), and PCLU is PCGrad applied inside the mask. FedCARE (arXiv 2601.22589) already does conflict-aware projection plus relearning-resistant recovery in FU, which is very close to PCLU + MCR. None of these are cited.

  6. The fusion rules in eq. 9-10 are only described in words, even in Appendix B, and the four deletion loss terms are never defined or weighted. The EMA statistic in eq. 13 isn't used anywhere. Eq. 8 uses the unprojected deletion gradient, and the code link is a placeholder. BN handling isn't stated either, and it matters here: local BN would give the model a domain id. Minor: hyphens are missing all over ("cross domain", "multi class"), there's "clents" in Fig. 1, and "Fluctuation Score" in Fig. 3 never shows up in the text.

Questions for authors
  1. What real scenario needs suppression beyond retrain, and how do you deal with the request being detectable?
  2. What is class c accuracy on each non-forget domain after unlearning?
  3. Does FA stay near 0 if FL keeps training for 20-50 more rounds?
  4. How were the hyperparameters for DeRe and the baselines picked, and how was FOUL adapted?
  5. Can you report std over seeds, the forget test set sizes, and results for other client/class pairs?

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

Related papers

Discussion (0)

Sign in to comment.