acceptodds
Under review as a conference paper at ICLR 2027

PDF: Path-Dependent Client Selection for Federated Unlearning

Abstract

Federated unlearning (FU) aims to remove target clients' influence from a trained model without full retraining. Existing methods largely treat unlearning as a post-hoc operation, overlooking that client influence is path-dependent: contributions made during critical learning periods (CLPs) can have disproportionate and persistent effects. Moreover, involving all retained clients is costly and impractical because clients differ in their usefulness for unlearning and may be intermittently available. We introduce PDF, a path-dependent client-selection framework that maintains lightweight, target-agnostic, CLP-weighted histories of each client's update direction and global-model alignment. Upon an unlearning request, PDF combines historical global alignment with target decoupling to select top- retained clients for retraining-based unlearning. Theoretically, we characterize the propagation of client perturbations, derive sufficient conditions for one-step retained-loss reduction and target-loss increase, and bound the initialization selection-utility gap assuming consistency between path-dependence scores and first-order client utility. Experiments on FMNIST, CIFAR-10, and CIFAR-100 under varying data heterogeneity in a label-flipping client-removal setting show that PDF achieves effective forgetting and strong retained utility using only approximately % of the retained clients. On non-IID CIFAR-10, PDF achieves % forgetting accuracy with an MIA AUC within percentage points of Retrain. On CIFAR-100, it improves retained accuracy by up to percentage points over the strongest evaluated FU baselines. Across evaluated scenarios, PDF reduces communication by approximately % and computation by -% relative to Retrain.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Client selection, Federated unlearning, Critical learning periods

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dbmPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PDF, a path-dependent forgetting method in federated learning. The paper is well-organized, the motivation is good. However, the impact of unlearning is not clearly validated in benign scenarios.

Critical strengths
  1. The idea is well-motivated and clearly stated. Though real practice has several gap with the theoretical analysis, the method is well-designed to align with the analysis.
  2. The method is lightweight and straight.
Critical weaknesses

I'm not quite familiar with Federated Unlearning (FU), but I have the following concerns:

  1. The experiments are mainly conducted in malicious scenarios where the FL framework are attacked by a label-flip-based attacker. Here, its quite confusing what is desired in such a scenario. General FU framework emphasize consistency with the retrained model. It would be helpful to clarify this distinction in metric choices as PDF treats lower FA and higher RA as better.

  2. If it is possible, please validate the PDF in benign FU scenarios. It is still not clear that if the model actually got rid of the historical influence of the target client, or if it mainly just fixed the performance bias caused by wrong labels.

  3. Label flip is a weak and outdated attack strategy. If the intention of malicious FU is to erase the influence of malicious clients, I think straight local tuning or continual federated training with same budget are also very effective.

Questions for authors

Please see weakness above. I'm willing to raise my score if main concerns are addressed.

8 comments
AuthorsOpenReview ↗

Response to Reviewer dbmP, part 1

We thank Reviewer for the constructive suggestions and comments. We address each question below.

Q1: The experiments are mainly conducted in malicious scenarios where the FL framework are attacked by a label-flip-based attacker. Here, its quite confusing what is desired in such a scenario. General FU framework emphasize consistency with the retrained model. It would be helpful to clarify this distinction in metric choices as PDF treats lower FA and higher RA as better.

a. The objective remains to approach retained-only retraining while preserving retained utility. Eq. (4) requires proximity to Retrain on the target objective and limits degradation on the retained objective. Accordingly, we have assessed FA by its absolute difference from Retrain, while retaining RA as a higher-is-better utility metric.

b. Role of label flipping. FA measures accuracy on the target’s held-out examples against the injected mapping y↦(y+1) mod Cy\mapsto(y+1)\bmod C, whereas RA uses retained clients’ correct labels. Label flipping makes agreement with a known target-associated mapping observable. Retrain can still have nonzero flipped-label FA through ordinary misclassification, despite never using the target’s data.

TABLE A. Label-flipping target results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. Retrain references train from scratch for 150 rounds, excluding the target. All remaining unlearning methods start from the same Original trained model, exclude the target, and perform 20 FedAvg rounds with the same local training. Continual FedAvg samples 16 retained clients per round, matching the original training protocol. Fresh-random selection resamples 25 retained clients each unlearning round, whereas fixed-random selection draws one 25-client subset and reuses it throughout all the unlearning rounds. PDF0.5\mathrm{PDF}_{0.5} uses the top 25 by Eq. (6), and PDF1.0\mathrm{PDF}_{1.0} sets β=1\beta=1, which switches off the target-decoupling term. Pooled MIA uses all 64 clients, as in the paper. Target-only MIA uses the removed client's training records as members and its held-out records as non-members.

MethodUnlearning client-roundsFA (%)RA (%)Pooled MIA AUC (%)Target-only MIA AUC (%)
Original (before removal)–7.38±4.317.38 \pm 4.3151.24±9.3451.24 \pm 9.3451.99±0.5251.99 \pm 0.5251.34±2.5751.34 \pm 2.57
Retrain16\mathrm{Retrain}_{16}, 16 per round24000.92±1.190.92 \pm 1.1946.05±7.3646.05 \pm 7.3650.70±2.0450.70 \pm 2.0449.31±3.9949.31 \pm 3.99
Retrain63\mathrm{Retrain}_{63}, all 63 per round94506.93±8.086.93 \pm 8.0855.28±4.5755.28 \pm 4.5755.64±1.0355.64\pm1.0352.67±7.6652.67\pm7.66
Continual FedAvg3203.89±5.923.89 \pm 5.9243.65±7.2143.65 \pm 7.2157.40±1.5357.40 \pm 1.5353.82±7.3353.82 \pm 7.33
Fresh-random5002.93±2.302.93 \pm 2.3052.58±13.6052.58 \pm 13.6057.98±1.7657.98 \pm 1.7652.14±7.4552.14 \pm 7.45
Fixed-random5004.58±6.334.58 \pm 6.3349.13±3.9449.13 \pm 3.9456.21±2.0056.21 \pm 2.0053.07±2.9153.07 \pm 2.91
PDF0.5\mathrm{PDF}_{0.5}5000.00±0.000.00 \pm 0.0054.33±7.54\mathbf{54.33 \pm 7.54}56.73±1.7356.73 \pm 1.7354.72±6.3554.72 \pm 6.35
PDF1.0\mathrm{PDF}_{1.0}5003.98±4.553.98 \pm 4.5559.98±7.08\mathbf{59.98 \pm 7.08}57.29±1.4557.29 \pm 1.4554.97±5.4854.97 \pm 5.48

Retrain16\mathrm{Retrain}_{16} follows Section 3.2 (from scratch, 150 rounds, 16 retained clients sampled per round) and is our reference below. Table 1 instead used all 63 retained clients per round, called Retrain63\mathrm{Retrain}_{63}, because it is the strongest reference on retained utility (RA 55.28% vs. 46.05%). It costs about four times higher communication overhead, so we report both for the label-flipping target.

c. Interpretation of the results. As shown in Table A, on CIFAR-10 with α=0.1\alpha=0.1, PDF0.5\mathrm{PDF}_{0.5} reduces flipped-label FA from 7.38% to 0.00% and increases RA from 51.24% to 54.33%. Its FA gaps are 0.92 percentage points to Retrain16\mathrm{Retrain}_{16} and 6.93 points to Retrain63\mathrm{Retrain}_{63}. Its RA is 8.28 points higher than Retrain16\mathrm{Retrain}_{16} and 0.95 points lower than Retrain63\mathrm{Retrain}_{63}. Thus, PDF\mathrm{PDF} achieves zero accuracy against the flipped labels while preserving retained accuracy near the full-participation Retrain level. We report the 6.93-point FA difference from that reference as a gap in target-data accuracy, rather than treating the lower FA as evidence of better unlearning.

Revision. We will replace the lower-is-better FA ranking with the smallest absolute FA gap to the specified Retrain reference, retain higher-is-better RA, and report pooled and target-specific MIA AUC with their retraining gaps. We assess these metrics jointly to characterize target-data accuracy, retained utility, and measured membership leakage. Results for correctly labelled targets (i.e., a benign setting) are provided in our response to Reviewer's Q2.

AuthorsOpenReview ↗

Response to Reviewer dbmP, part 2

Q2: If it is possible, please validate the PDF in benign FU scenarios. It is still not clear that if the model actually got rid of the historical influence of the target client, or if it mainly just fixed the performance bias caused by wrong labels.

As Reviewer suggested, we have conducted experiments on a benign setting, where we try to unlearn a correctly labelled client. Table B below includes new results.

TABLE B. Benign target results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. FA is the accuracy on the target's own held-out data, whose labels are correct.

MethodParticipantsUnlearning client-roundsFA (%)RA (%)Target-only MIA AUC (%)
Original (before removal)16 of 64 per round, target included, 150 rounds–50.33±28.3350.33 \pm 28.3352.94±3.8752.94 \pm 3.8751.82±2.00\mathbf{51.82 \pm 2.00}
Retrain16 retained per round, 150 rounds from scratch240059.88±32.3859.88 \pm 32.3846.06±12.0546.06 \pm 12.0547.82±1.3547.82 \pm 1.35
Continual FedAvg16 retained, redrawn each round, 20 rounds32032.63±16.3332.63 \pm 16.3346.17±4.4946.17 \pm 4.4950.53±3.7850.53 \pm 3.78
Fresh-random25 retained, redrawn each round, 20 rounds50044.28±25.5044.28 \pm 25.5051.35±10.8051.35 \pm 10.8052.57±5.4152.57 \pm 5.41
Fixed-random25 retained, drawn once, 20 rounds50042.46±13.7742.46 \pm 13.7747.32±3.7347.32 \pm 3.7352.02±3.5852.02 \pm 3.58
PDF0.5\mathrm{PDF}_{0.5}Top 25 by Eq. (6), 20 rounds50022.99±23.0122.99 \pm 23.0152.87±6.30\mathbf{52.87 \pm 6.30}50.67±2.9250.67 \pm 2.92
PDF1.0\mathrm{PDF}_{1.0}Top 25 by Eq. (6) with β=1\beta=1, 20 rounds50050.60±19.5350.60 \pm 19.5359.21±5.9759.21 \pm 5.9752.16±6.0952.16 \pm 6.09

a. The effect is not only wrong label repair. With the same unlearning budget, PDF0.5\mathrm{PDF}_{0.5} achieves 52.87% RA, compared with 47.32% and 51.35% for fixed-random and fresh-random selection. It also preserves the original trained model’s 52.94% RA. These results demonstrate that historical client selection provides a retained-utility benefit even when there are no corrupted labels to correct.

b. Target behavior and shared knowledge. Retrain achieves 59.88% target accuracy using retained clients alone. PDF0.5\mathrm{PDF}_{0.5} reaches 22.99%, indicating excessive suppression of benign-target performance. Removing target decoupling (β=1\beta=1) raises target accuracy to 50.60% and RA to 59.21%, reducing the target-accuracy gap to Retrain from 36.89 to 9.28 points. The resulting target accuracy is also near the original model’s 50.33%, showing that this ablation primarily preserves shared-class performance. Therefore, this shows that PDF\mathrm{PDF} provides flexibility through β\beta, which balances historical global alignment and target decoupling tailored for each task (e.g., increasing β\beta for benign settings).

c. Target-specific membership inference. For PDF0.5\mathrm{PDF}_{0.5}, mean target-specific AUC decreases from 51.82% to 50.67%, narrowing its gap to Retrain (i.e., 47.82%) from 4.00 to 2.85 percentage points. With β=1\beta=1, AUC is 52.16%, showing that the gains in RA and FA accompany higher membership distinguishability.

These experiments establish that PDF\mathrm{PDF}'s utility benefit extends beyond label correction and clarify how target decoupling affects benign removal.

AuthorsOpenReview ↗

Response to Reviewer dbmP, part 3

Q3: Label flip is a weak and outdated attack strategy. If the intention of malicious FU is to erase the influence of malicious clients, I think straight local tuning or continual federated training with same budget are also very effective.

As clarified in Q1, we use label flipping as a controlled diagnostic of agreement with an injected label mapping. Our focus is retained-client selection for unlearning. Continual federated training provides a direct baseline because PDF\mathrm{PDF} continues FedAvg from the trained model and changes which retained clients participate through Eq. (6). Since the server holds no training data, we interpret straight local tuning as fine-tuning on retained clients’ local data using the same FedAvg protocol, with random participation instead of PDF\mathrm{PDF}'s CLP-weighted selection.

We agree with Reviewer that continual training without selection is also effective, as shown in Table A. Particularly, they lower FA from 7.38% to between 2.93% (i.e., fresh-random) and 4.58% (i.e., fixed-random), which removes 38% to 60% of the flipped-label behaviour. Besides, Continual FedAvg with 16 sampled clients per round achieves 3.89% FA using 320 participations. Therefore, PDF\mathrm{PDF} contributes two things:

  • Target decoupling completes the removal. PDF0.5\mathrm{PDF}_{0.5} reaches 0.00% FA across all four seeds, compared with 3.98% for PDF1.0\mathrm{PDF}_{1.0} without target decoupling, under the same participation budget and training protocol. This ablation directly demonstrates the term’s contribution to suppressing flipped-label predictions. Relative to Retrain16\mathrm{Retrain}_{16}, which achieves 0.92% FA, the absolute FA gap decreases from 3.06 to 0.92 points.

  • Historical global alignment preserves utility. At the same 500 client participations, mean RA is 54.33% for PDF0.5\mathrm{PDF}_{0.5} and 59.98% for PDF1.0\mathrm{PDF}_{1.0}, compared with 49.13% for fixed-random and 52.58% for fresh-random selection. Both PDF variants also exceed the original model’s 51.24% RA. The β=1\beta=1 result shows that selection based on historical global alignment remains useful without target decoupling, while β\beta controls the balance between retained utility and suppression of the injected mapping.

Revision. We will add the benign setting and the three same-budget fine-tuning baselines to Section 5.3, and discuss the role of β\beta for benign and label-flipped targets in Sections 4.2 and 5.3.

Reviewer_k8jAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PDF, a client-selection method for federated unlearning. Its central idea is that a retained client’s usefulness for unlearning depends on its contributions during the original training.

During training, PDF stores a compact history of each client’s classification-layer update direction and its alignment with the global update. Contributions during detected critical learning periods receive greater weight. When a removal request arrives, PDF selects retained clients whose histories align with the global trajectory but differ from the target client’s history. It then continues standard federated training from the existing model with only the selected clients.

The paper provides perturbation-propagation bounds, one-step loss bounds, and a conditional guarantee for client selection. Experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 evaluate the removal of a label-flipping client. PDF achieves competitive retained accuracy using 25 of the 63 retained clients, and the paper reports communication and computation savings relative to retraining.

Critical strengths
  1. The method addresses a practical selection problem with modest historical storage. PDF asks which retained clients are most useful after a removal request. Its histories are constructed before the target is known, and storage does not grow with the number of training rounds. Storing one classification-layer vector and two scalars per client is a practical design choice.

The method also preserves standard local optimization and aggregation. Its contribution is therefore easy to isolate, and it could be combined with other unlearning methods.

  1. The paper provides supporting diagnostics beyond the main performance table. Appendix A.4.1 reruns training after omitting a single client contribution while keeping subsequent participation and randomness coupled. This is a useful way to test whether contributions made during detected critical periods have larger downstream effects.

The paper also measures the relationship between historical scores and initialization utility. The reported Spearman correlation of approximately 0.59 provides evidence that the score contains useful information, although it is an imperfect proxy. These diagnostics give the motivation empirical support beyond an intuition about training history.

Critical weaknesses
  1. The evaluation does not distinguish client unlearning from correction of corrupted labels.

All reported main experiments remove a label-flipping client, and forgetting accuracy is measured against its flipped labels. A model that learns the correct classification rule can therefore achieve very low forgetting accuracy while still retaining information learned from that client’s images.

A model can stop predicting the malicious label for an image without losing the influence that image had on its representation. Zero accuracy on flipped labels is therefore not sufficient evidence of successful client removal.

This also creates a mismatch with Equation (4), which seeks retraining-like behavior. On CIFAR-10 with heterogeneity parameter 0.1, Retrain has forgetting accuracy of 6.93%, while PDF achieves 0%. The paper presents the 0% as stronger forgetting, but this metric alone does not establish that PDF is closer to the retrained model.

The membership-inference evaluation could provide complementary evidence. However, Appendix A.5.1 describes querying “original FL training samples” without specifying whether evaluation is restricted to the removed client. If retained-client samples are included, aggregate AUC may obscure residual leakage from the target. The main table also omits the original, unchanged model, so it is unclear how much target-specific leakage existed before unlearning.

This concern matters because the paper’s central claim is removal of client influence, whereas the current evidence more directly shows recovery from a particular poisoning scenario.

  1. The experiments do not isolate the benefit of the proposed selection rule under comparable budgets.

PDF’s unlearning operation is standard fine-tuning on selected retained clients. The essential comparison is therefore against the same fine-tuning procedure with randomly selected clients and identical initialization, client count, rounds, and local computation. That comparison is absent.

Table 3 compares PDF with reversed ranking, high target coherence, and removal of critical-period weighting. These are useful ablations, but reversed ranking and high coherence deliberately select clients that the proposed hypothesis considers undesirable, so beating them does not show that PDF improves over ordinary random participation. The non-CLP variant still uses the paper’s other selection components.

The efficiency comparison also needs stronger control. Original training samples 16 clients per round, whereas the discussion describes Retrain as involving all retained clients. Retraining on the retained population does not require every retained client to participate in every round. If the schedules differ, some of the reported savings may arise from the baseline configuration rather than from the proposed selection method. Separate training and unlearning costs, together with matched participation schedules, are needed.

The theoretical guarantee does not close this gap. Corollary 1 assumes that the historical score already approximates initialization utility and then bounds the resulting selection error. It does not establish that PDF satisfies a small error bound. The reported normalized maximum approximation error is approximately 0.4309, so the corresponding average-utility guarantee is 2ϵ≈0.8618 range⁡(Q)2\epsilon \approx 0.8618\,\operatorname{range}(Q), which allows a gap approaching the full observed utility range. The measured selection performance is more encouraging than this bound, but it is evaluated in one Fashion-MNIST setting and does not replace the missing matched-budget comparisons.

Questions for authors
  1. Forgetting evaluation: Is MIA AUC computed only on the removed client’s training samples and corresponding nonmember samples? Please report target-specific attack results for the original model, Retrain, and PDF, with uncertainty across runs. This would show whether the attack detects leakage before removal and whether PDF reduces it.
  2. Selection benefit: How does PDF compare with a fixed random subset of 25 retained clients and with fresh random participation each round, given the same starting model, the same 20 rounds, and the same local computation? These comparisons would directly test the paper’s main contribution.
  3. Retraining reference and cost: How many clients participate in each Retrain round, and does the Retrain cost include the original training stage? Please provide separate training and unlearning costs and a retraining comparison using the original 16-client sampling protocol.
  4. General client removal: Does the method approach retraining when the removed client has correct labels and shares classes with retained clients? In this setting, increasing target loss may suppress legitimate shared knowledge. The results would show whether target decoupling supports general client removal or primarily malicious-client repair.
7 comments
AuthorsOpenReview ↗

Response to Reviewer k8jA, part 1

We thank Reviewer for the constructive suggestions and comments. We address each weakness (W) and question (Q) below.

W1: The evaluation does not convincingly distinguish client unlearning from correction of corrupted labels. All reported main experiments remove a label-flipping client, and forgetting accuracy is measured against its flipped labels... Consequently, zero accuracy on flipped labels is not sufficient evidence of successful client removal. This also creates a mismatch with Equation (4), which seeks retraining-like behavior. On CIFAR-10 with heterogeneity parameter 0.1, Retrain has forgetting accuracy of 6.93%, while PDF achieves 0%. The latter is presented as stronger forgetting, but this metric alone does not establish that PDF is closer to the retrained model. The membership-inference evaluation could provide complementary evidence, but Appendix A.5.1 describes querying “original FL training samples” without clearly specifying whether evaluation is restricted to the removed client. If retained-client samples are included, aggregate AUC may obscure residual leakage from the target. The main table also omits the original, unchanged model, so it is unclear how much target-specific leakage existed before unlearning. This concern is consequential because the central claim is removal of client influence...

Q4: General client removal: Does the method approach retraining when the removed client has correct labels and shares classes with retained clients? In this setting, increasing target loss may suppress legitimate shared knowledge. Results would clarify whether target decoupling supports general client removal or primarily malicious-client repair.

We have added experiments with label-flipping and benign targets, including the original trained model, target-specific MIA, and a β=1\beta=1 ablation that removes target decoupling. Tables A and B report results over four independent random seeds on CIFAR-10 with α=0.1\alpha=0.1.

TABLE A. Label-flipping target results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. Retrain references train from scratch for 150 rounds, excluding the target. All remaining unlearning methods start from the same Original trained model, exclude the target, and perform 20 FedAvg rounds with the same local training. Continual FedAvg samples 16 retained clients per round, matching the original training protocol. Fresh-random selection resamples 25 retained clients each unlearning round, whereas fixed-random selection draws one 25-client subset and reuses it throughout all the unlearning rounds. PDF\mathrm{PDF} variants across β∈0.1,0.5,0.9,1.0\beta \in \\{0.1, 0.5, 0.9, 1.0\\} use the top 25 by Eq. (6), and PDF1.0\mathrm{PDF}_{1.0} sets β=1\beta=1, which switches off the target-decoupling term. Pooled MIA uses all 64 clients, as in the paper. Target-only Shokri MIA and Target-only Loss MIA (i.e., per-record loss, no attack model) use the removed client's training records as members and its held-out records as non-members.

MethodUnlearning client-roundsFA (%)RA (%)Pooled MIA AUC (%)Target-only Shokri MIA (%)Target-only Loss MIA (%)
Original (before removal)–7.38±4.317.38 \pm 4.3151.24±9.3451.24 \pm 9.3451.99±0.5251.99 \pm 0.5251.34±2.5751.34 \pm 2.5751.67±5.9651.67 \pm 5.96
Retrain16\mathrm{Retrain}_{16}, 16 per round24000.92±1.190.92 \pm 1.1946.05±7.3646.05 \pm 7.3650.70±2.0450.70 \pm 2.0449.31±3.9949.31 \pm 3.9950.27±2.1750.27 \pm 2.17
Retrain63\mathrm{Retrain}_{63}, all 63 per round94506.93±8.086.93 \pm 8.0855.28±4.5755.28 \pm 4.5755.64±1.0355.64 \pm 1.0352.67±7.6652.67 \pm 7.6651.58±4.3151.58 \pm 4.31
Continual FedAvg3203.89±5.923.89 \pm 5.9243.65±7.2143.65 \pm 7.2157.40±1.5357.40 \pm 1.5353.82±7.3353.82 \pm 7.3351.60±6.8651.60 \pm 6.86
Fresh-random5002.93±2.302.93 \pm 2.3052.58±13.6052.58 \pm 13.6057.98±1.7657.98 \pm 1.7652.14±7.4552.14 \pm 7.4550.04±5.1650.04 \pm 5.16
Fixed-random5004.58±6.334.58 \pm 6.3349.13±3.9449.13 \pm 3.9456.21±2.0056.21 \pm 2.0053.07±2.9153.07 \pm 2.9149.67±1.0849.67 \pm 1.08
PDF0.1\mathrm{PDF}_{0.1}5001.98±2.951.98 \pm 2.9543.30±5.8443.30 \pm 5.8455.13±1.2755.13 \pm 1.2750.88±2.9950.88 \pm 2.9950.52±2.2050.52 \pm 2.20
PDF0.5\mathrm{PDF}_{0.5}5000.00±0.000.00 \pm 0.0054.33±7.5454.33 \pm 7.5456.73±1.7356.73 \pm 1.7354.72±6.3554.72 \pm 6.3549.99±2.6749.99 \pm 2.67
PDF0.9\mathrm{PDF}_{0.9}5000.86±0.560.86 \pm 0.5655.16±5.6455.16 \pm 5.6456.14±0.8456.14 \pm 0.8453.95±5.0653.95 \pm 5.0649.79±1.6649.79 \pm 1.66
PDF1.0\mathrm{PDF}_{1.0}5003.98±4.553.98 \pm 4.5559.98±7.0859.98 \pm 7.0857.29±1.4557.29 \pm 1.4554.97±5.4854.97 \pm 5.4849.59±3.3949.59 \pm 3.39

Retrain16\mathrm{Retrain}_{16} follows Section 3.2 (from scratch, 150 rounds, 16 retained clients sampled per round) and is our reference below. Table 1 instead used all 63 retained clients per round, called Retrain63\mathrm{Retrain}_{63}, because it is the strongest reference on retained utility (RA 55.28% vs. 46.05%). It costs about four times higher communication overhead, so we report both for the label-flipping target.

AuthorsOpenReview ↗

Response to Reviewer k8jA, part 2

(Continue to respond to W1 and Q4)

TABLE B. Benign target results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. FA is the accuracy on the target's own held-out data, whose labels are correct.

MethodParticipantsUnlearning client-roundsFA (%)RA (%)Target-only Shokri MIA (%)Target-only Loss MIA (%)
Original (before removal)16 of 64 per round, target included, 150 rounds–50.33±28.3350.33 \pm 28.3352.94±3.8752.94 \pm 3.8751.82±2.0051.82 \pm 2.0050.55±3.5650.55 \pm 3.56
Retrain16\mathrm{Retrain}_{16}16 retained per round, 150 rounds from scratch240059.88±32.3859.88 \pm 32.3846.06±12.0546.06 \pm 12.0547.82±1.3547.82 \pm 1.3548.37±5.8848.37 \pm 5.88
Continual FedAvg16 retained, redrawn each round, 20 rounds32032.63±16.3332.63 \pm 16.3346.17±4.4946.17 \pm 4.4950.53±3.7850.53 \pm 3.7849.69±2.8649.69 \pm 2.86
Fresh-random25 retained, redrawn each round, 20 rounds50044.28±25.5044.28 \pm 25.5051.35±10.8051.35 \pm 10.8052.57±5.4152.57 \pm 5.4149.05±4.1249.05 \pm 4.12
Fixed-random25 retained, drawn once, 20 rounds50042.46±13.7742.46 \pm 13.7747.32±3.7347.32 \pm 3.7352.02±3.5852.02 \pm 3.5849.81±5.0049.81 \pm 5.00
PDF0.1\mathrm{PDF}_{0.1}Top 25 by Eq. (6) with β=0.1\beta=0.1, 20 rounds50019.06±13.7819.06 \pm 13.7844.71±12.2744.71 \pm 12.2748.79±5.3848.79 \pm 5.3848.34±4.8348.34 \pm 4.83
PDF0.5\mathrm{PDF}_{0.5}Top 25 by Eq. (6) with β=0.5\beta=0.5, 20 rounds50022.99±23.0122.99 \pm 23.0152.87±6.30\mathbf{52.87 \pm 6.30}50.67±2.9250.67 \pm 2.9248.67±5.1548.67 \pm 5.15
PDF0.9\mathrm{PDF}_{0.9}Top 25 by Eq. (6) with β=0.9\beta=0.9, 20 rounds50044.45±31.1744.45 \pm 31.1757.04±2.3757.04 \pm 2.3753.12±3.7853.12 \pm 3.7847.30±4.5847.30 \pm 4.58
PDF1.0\mathrm{PDF}_{1.0}Top 25 by Eq. (6) with β=1\beta=1, 20 rounds50050.60±19.5350.60 \pm 19.5359.21±5.9759.21 \pm 5.9752.16±6.0952.16 \pm 6.0948.50±4.5648.50 \pm 4.56

a. FA relative to Retrain (Eq. 4). PDF\mathrm{PDF} reduces flipped-label FA from 7.38% to 0.00%, while increasing RA from 51.24% to 54.33%. This demonstrates suppression of the injected mapping and improved RA. To align the interpretation with Eq. (4), we will assess FA by its absolute gap to Retrain: PDF’s gaps are 0.92 points to Retrain16\mathrm{Retrain}_{16} and 6.93 points to Retrain63\mathrm{Retrain}_{63}.

b. Benign removal and shared knowledge. With correct target labels, PDF0.5\mathrm{PDF}_{0.5} preserves RA at 52.87%, compared with 52.94% before removal, demonstrating utility preservation beyond corrupted-label correction. Although its FA is 22.99%, versus 59.88% for Retrain, removing target decoupling (i.e., β=1\beta=1) raises RA to 59.21% and FA to 50.60%, which reduces the FA gap from 36.89 to 9.28 percentage points. This identifies the contribution of target decoupling to suppressing shared-class performance. For the label-flipping target, the same ablation raises flipped-label FA from 0.00% to 3.98%, showing its role in suppressing the injected mapping.

c. Membership inference. The manuscript’s pooled AUC includes all 64 clients. The added target-specific evaluation uses only the removed client’s training and held-out records. Under label flipping, PDF0.5\mathrm{PDF}_{0.5}'s target-specific Shokri AUC is 54.72%, versus 51.34% before removal, 49.31% for Retrain16\mathrm{Retrain}_{16}, and 52.67% for Retrain63\mathrm{Retrain}_{63}. Under benign removal, it decreases from 51.82% to 50.67%, compared with 47.82% for Retrain, setting β=1\beta=1 gives 52.16%. To further separate instance-level memorization from changes in prediction confidence as RA improves, we also report Target-only Loss MIA (last column of Tables A and B), which directly measures whether per-sample loss separates the target client's training records from its held-out records without an auxiliary shadow classifier. Under label flipping (Table A), Target-only Loss MIA decreases from 51.67% (Original) to 49.59%–50.52% across all PDF\mathrm{PDF} variants, matching Retrain16\mathrm{Retrain}_{16} (50.27%) within 0.25–0.68 points. Under benign removal (Table B), Target-only Loss MIA drops from 50.55% (Original) to 47.30%-48.67% across all PDF\mathrm{PDF} variants, including PDF0.9\mathrm{PDF}_{0.9} (47.30%) and PDF1.0\mathrm{PDF}_{1.0} (48.50%), aligning with Retrain16\mathrm{Retrain}_{16} (48.37%, within 0.03–1.07 points) and achieving closer agreement to Retrain16\mathrm{Retrain}_{16} than Continual FedAvg (49.69%), Fresh-random (49.05%), and Fixed-random (49.81%).

These results demonstrate PDF\mathrm{PDF}'s RA benefits and show how adjusting target decoupling improves the balance between preserving shared knowledge and approaching Retrain's FA in both label-flipping and benign settings.

Revision. Table 1 will report the Original model. Section 5 will add the benign setting, the role of β\beta, and the target-specific MIA. Besides, Appendix A.5.1 will specify the cohort used for the pooled AUC.

AuthorsOpenReview ↗

Response to Reviewer k8jA, part 3

W2: The experiments do not isolate the benefit of the proposed selection rule under comparable budgets. The theoretical guarantee does not close this gap. Corollary 1 assumes that the historical score already approximates initialization utility and then bounds the resulting selection error. It does not establish that PDF satisfies a small error bound.Indeed, the reported normalized maximum approximation error is approximately 0.4309, making the corresponding average-utility guarantee 2ϵ≈0.8618,range(Q)2\epsilon \approx 0.8618, \text{range}(Q). This allows a gap approaching the full observed utility range. The measured selection performance is more encouraging than this bound, but it is evaluated in one Fashion-MNIST setting and does not replace the missing matched-budget comparisons.

First, to empirically isolate the selection rule under matched budgets, we have added comparisons against fixed-random and fresh-random 25-client selection in Tables A and B above (detailed in Q2 below). Regarding the theoretical guarantee, Corollary 1 is a conditional result, as we mentioned in its title, the abstract, and Section 4.3 indicate. It holds for any ϵ≥0\epsilon\ge0 and characterizes how selection quality depends on PD-utility consistency. It does not claim that PD attains a small ϵ\epsilon, which is why Appendix A.4.3 on Page 18 measures ϵ\epsilon and the selection quality directly.

The figure 2ϵ≈0.8618 range(Q)2\epsilon\approx0.8618\,\mathrm{range}(Q) is correct, and it is a worst-case constant. ϵaff\epsilon_{\mathrm{aff}} is the maximum absolute residual over all retained clients, so it is set by the worst-fitted client and then applied to every selected client. Appendix A.4.3 already describes the resulting guarantee as valid but conservative: the realized gap (0.32180.3218) is 8.998.99\\% of the bound (i.e., about 0.08 range(Q)0.08\,\mathrm{range}(Q) on average rather than 0.860.86).

The figure 2ϵ≈0.8618 range(Q)2\epsilon\approx0.8618\,\mathrm{range}(Q) is correct, and it is a worst-case constant. ϵaff\epsilon_{\mathrm{aff}} is the maximum absolute residual over all retained clients, so it is set by the worst-fitted client and then applied to every selected client. Appendix A.4.3 already describes the resulting guarantee as valid but conservative: the realized gap (0.32180.3218) is 8.998.99\\% of the bound (i.e., about 0.08 range(Q)0.08\,\mathrm{range}(Q) on average rather than 0.860.86).

The paper's evidence for selection quality is therefore the measurement, not the constant. At the same initialization, PD agrees with QiQ_i in ranking (Spearman 0.590.59, 95% CI [0.48,0.71][0.48,0.71], pairwise accuracy 71.871.8\\%), overlaps the utility-optimal subset K⋆\mathcal K^\star by 7070\\% (chance 25/63≈4025/63\approx40\\%), and attains 7777\\% of the optimal summed utility (1.10411.1041 vs. 1.42591.4259). The corollary's role is to identify the quantity that must be checked and Table 6 checks it.

Revision. We will move the measured ϵ\epsilon to Section 4.3.


Q1: Is MIA AUC computed exclusively on the removed client's training samples and corresponding nonmember samples? Please report target-specific attack results for the original model, Retrain, and PDF, with uncertainty across seeds.

No. The MIA AUC in the paper is computed on a pooled cohort, not on the target client alone as we responded in W1(c) above. As Reviewer suggested, we have run the same attack with members and nonmembers restricted to the removed client's training and held-out records (last two columns of Tables A and B, reporting target-only Shokri and loss-based attacks).

a. How the reported AUC is computed. Our MIA AUC has been computed on a pooled membership cohort, not exclusively on the removed client's samples. We follow the shadow-model attack of Shokri et al. [1]: five shadow unlearning runs are performed on the proxy splits, which are disjoint from the clients' training and evaluation data. A three-layer MLP attack classifier is trained on the shadow models' softmax outputs, with proxy-train records labeled as members and proxy-evaluation records labeled as nonmembers. At evaluation time, members are client training records and nonmembers are held-out client evaluation records, both pooled over all 64 clients and balanced to equal size. The target client therefore contributes only about 1/64 of the queries.

b.1. Does the attack detect target leakage before removal? The Original model’s target-only Shokri MIA AUC is 51.82±2.0051.82 \pm 2.00\\% for the benign target and 51.34±2.5751.34 \pm 2.57\\% for the label-flipping target, respectively 1.82 and 1.34 percentage points above chance. These mean results indicate limited membership discrimination before removal. Compared with Retrain16\mathrm{Retrain}_{16}, which excludes the target and follows the original sampling protocol, the Original model’s AUC is 4.00 points higher in the benign setting and 2.03 points higher under label flipping.

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

This paper studies client-level federated unlearning and proposes PDF, a method for choosing which retained clients take part in retraining-based unlearning. During FedAvg training, the server detects critical learning periods, or CLPs, from the relative change of the federated gradient norm, or FGN. For each client, it keeps two statistics on the classification layer, both weighted more heavily in CLP rounds. One is the accumulated update direction, and the other is how well the client's updates agreed with the global update. This needs one classifier-sized vector and two scalars per client, and it does not depend on any future request. When a request arrives, each retained client gets a score with two parts. One is its past agreement with the global update, weighted by β. The other is how far its past direction is from that of the target clients, which gets the remaining weight. The top k retained clients then run 20 FedAvg rounds starting from the trained model, and the target clients do not take part.

Critical strengths
  1. The idea is simple and cheap, and the results are promising. Choosing retained clients from histories collected during training is separate from the unlearning algorithm itself, and the history costs only one classifier-sized vector and two scalars per client. On CIFAR-100 all three PDF variants beat the best competing method in RA while keeping FA below 1%. They reach 48.43 to 49.13 against 42.22 for FOUL at \alpha = 0.1, and 54.21 to 54.38 against 47.50 for GDFA at \alpha = 1.0. The ablations also show that the ranking matters. On CIFAR-10 with \alpha = 0.1, reversing the ranking gives 41.09 RA and choosing clients close to the target gives 37.65, compared with 54.48 for PDF under the same unlearning procedure. This is the main reason I think the idea is worth pursuing.

  2. The paper tests its key assumptions directly and is open about weaker results. In Appendix A.4 the authors drop one client from a single round, rerun the remaining rounds with the same randomness, and repeat this over six seeds. Contributions made in CLP rounds shift the final parameters by 1.1356 on average, and other contributions by 0.7176, and the 95% confidence interval of the gap is 0.2915 to 0.5444. The score also ranks clients in line with their first-order utility, with a Spearman correlation of 0.5934, and the top 25 clients share 0.7 of their members with the best subset on average. A random subset of 25 out of 63 clients would share about 0.40 in expectation, so this overlap is well above chance. These studies suggest that the score carries real signal, but they only use FMNIST with \alpha = 0.1. The paper also reports an approximation error of 0.4309 relative to the range of utilities, and a gap of 9.12 to 9.76 points in MIA AUC to Retrain on CIFAR-100 with α = 1.0.

Critical weaknesses
  1. The main claim is not separated from simpler explanations. The claim is that training histories pick a better set of retained clients. Since PDF only changes who takes part in 20 rounds of fine-tuning from the trained model, three comparisons are needed. The first is a random choice of 25 retained clients with the same number of rounds. The second is plain fine-tuning with all retained clients, or with 16 sampled per round as in training, at the same number of rounds or the same total client rounds. The third is the original model before unlearning, which shows where FA, RA, and MIA start. None of them is reported. Reversing the ranking and choosing clients close to the target are useful ablations, but they work against the method by design and cannot stand in for a neutral baseline. Several baselines have far lower RA than Retrain. This weakness matters most for my rating, because without these controls I cannot tell how much of the gain comes from the score and how much comes from fine-tuning on a subset of clean clients.

  2. The experiments cover a narrow setting, and the theory gives little support for the specific design. All reported unlearning experiments remove one client with flipped labels out of 64. The method is defined for several targets, but only one is ever tested. The paper also does not try removing a benign client, which is the GDPR case stressed in the introduction, or handling requests one after another. FA is measured on the flipped labels, and lower FA always counts as better, even below Retrain.

  3. Proposition 1 is a standard bound obtained by unrolling Lipschitz steps. Its weights \omega_t never grow with t, because each factor \Gamma_s is at least 1. The bound does not involve FGN and does not show that timing changes the actual influence. The support for CLP weighting therefore comes from the ablation without CLP weighting and from the FMNIST counterfactual study, and that study does not compare rounds at the same stage of training. Proposition 2 is the standard descent lemma. Corollary 1 follows directly from Assumption 1, and the paper assumes rather than derives that the error \epsilon is small for its score.

Questions for authors
  1. In Table 1, Retrain communication divided by CONDA communication is about 3.94 on all three datasets, namely 15.62 over 3.97, 845.48 over 214.73 and 848.97 over 215.61. This is close to 63/16. PDF uses 44.73 GB more than CONDA on CIFAR-10, which matches about 500 extra client rounds, or 25 clients for 20 rounds, against 2400 client rounds in training. With a fixed cost per client round, the Retrain numbers look like one run of 150 rounds with all 63 retained clients, compared with a training run of 16 clients per round. If training were also counted for Retrain, as Section 5.1 says it is, the ratio would be close to 4.94. Please report training and unlearning communication separately, and state the number of rounds and clients per round used by Retrain. Does Retrain sample clients the same way as the original training, as Section 3.2 defines it? With 16 clients per round for 150 rounds and training counted for both methods, the saving would be about 40% rather than 69%.
7 comments
AuthorsOpenReview ↗

Response to Reviewer giuR, part 1

We thank Reviewer for your suggestions and comments. We would like to address and clarify your concerns about weaknesses (W) and questions (Q) as follows.

W1: The main claim is not separated from simpler explanations... The first is a random choice of 25 retained clients with the same number of rounds. The second is plain fine-tuning with all retained clients, or with 16 sampled per round as in training, at the same number of rounds or the same total client rounds. The third is the original trained model before unlearning, which shows where FA, RA, and MIA start... I cannot tell how much of the gain comes from the score and how much comes from fine-tuning on a subset of clean clients.

Based on Reviewer's suggestions, we have implemented new ablations and reported results over 4 independent random seeds in Table A below with corresponding analysis.

TABLE A. Label-flipping target results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. Retrain references train from scratch for 150 rounds, excluding the target. All remaining unlearning methods start from the same Original trained model, exclude the target, and perform 20 FedAvg rounds with the same local training. Continual FedAvg samples 16 retained clients per round, matching the original training protocol. Fresh-random selection resamples 25 retained clients each unlearning round, whereas fixed-random selection draws one 25-client subset and reuses it throughout all the unlearning rounds. PDF0.5\mathrm{PDF}_{0.5} uses the top 25 by Eq. (6). Pooled MIA uses all 64 clients, as in the paper. Target-only MIA uses the removed client's training records as members and its held-out records as non-members.

MethodRetained clients neededUnlearning client participationsFA (%)RA (%)MIA AUC (%)Target-specific MIA AUC (%)
Original FL model (before unlearning)–07.38±4.31\mathbf{7.38 \pm 4.31}51.24±9.3451.24 \pm 9.3451.99±0.5251.99 \pm 0.5251.34±2.5751.34 \pm 2.57
Retrain16\mathrm{Retrain}_{16}, 16 per roundup to 6324000.92±1.190.92 \pm 1.1946.05±7.3646.05 \pm 7.3650.70±2.0450.70 \pm 2.0449.31±3.9949.31 \pm 3.99
Retrain63\mathrm{Retrain}_{63}, all 63 per round6394506.93±8.086.93 \pm 8.0855.28±4.5755.28 \pm 4.5755.64±1.0355.64\pm1.0352.67±7.6652.67\pm7.66
Continual FedAvg, 16 sampled per roundup to 633203.89±5.923.89 \pm 5.9243.65±7.21\mathbf{43.65 \pm 7.21}57.40±1.5357.40 \pm 1.5353.82±7.3353.82 \pm 7.33
Fresh-random 25up to 635002.93±2.302.93 \pm 2.3052.58±13.6052.58 \pm 13.6057.98±1.7657.98 \pm 1.7652.14±7.4552.14 \pm 7.45
Fixed-random 25255004.58±6.334.58 \pm 6.3349.13±3.9449.13 \pm 3.9456.21±2.0056.21 \pm 2.0053.07±2.9153.07 \pm 2.91
PDF0.5\mathrm{PDF}_{0.5}, top-25 fixed255000.00±0.00\mathbf{0.00 \pm 0.00}54.33±7.54\mathbf{54.33 \pm 7.54}56.73±1.7356.73 \pm 1.7354.72±6.3554.72 \pm 6.35

a. Improvement over the original model. PDF\mathrm{PDF} increases RA from 51.24% to 54.33% while reducing flipped-label FA from 7.38% to 0.00%, using a fixed subset of only 25 of 63 retained clients. These results demonstrate the contribution of historical selection to preserving retained utility while suppressing the injected label mapping.

b. CLP-weighted selection benefit at the same budget. With the same checkpoint, 20 unlearning rounds, and 500 client participations, PDF\mathrm{PDF} reduces FA to 0.00% from 4.58% and 2.93% for fixed-random and fresh-random selection, while increasing mean RA by 5.20 and 1.75 percentage points. The fixed-random comparison is particularly informative: both methods reuse 25 clients, so the accuracy gains support the value of which clients PDF\mathrm{PDF} selects.

c. Benefit over continual FedAvg. PDF\mathrm{PDF} achieves 54.33% RA and 0.00% FA, versus 43.65% RA and 3.89% FA when continuing the original 16-client sampling protocol.

d. MIA trade-offs. PDF\mathrm{PDF}'s pooled and target-specific MIA AUC are 56.73% and 54.72%, compared with 51.99% and 51.34% before unlearning. Relative to fixed-random, fresh-random, and continual FedAvg, its pooled AUC differences are +0.52, -1.25, and -0.67 percentage points, while its target-specific differences are +1.65, +2.58, and +0.90 points. For comparison, Retrain16\mathrm{Retrain}_{16} achieves 50.70% pooled AUC and 49.31% target-specific AUC, whereas Retrain63\mathrm{Retrain}_{63} achieves 55.64% and 52.67%, respectively. PDF’s corresponding gaps are +6.03 and +5.41 points against Retrain16\mathrm{Retrain}_{16}, and +1.09 and +2.05 points against Retrain63\mathrm{Retrain}_{63}.

Overall, PDF\mathrm{PDF} improves RA and suppresses flipped-label predictions while achieving reasonable MIA performance in this setting, with a mean target-specific AUC of 54.72%. These results support the value of path-dependent client selection in balancing retained utility, forgetting behavior, and membership privacy.

AuthorsOpenReview ↗

Response to Reviewer giuR, part 2

W2: ...All reported unlearning experiments remove one client with flipped labels out of 64... The paper also does not try removing a benign client, which is the GDPR case stressed in the introduction, or handling requests one after another. FA is measured on the flipped labels, and lower FA always counts as better, even below Retrain.

As Reviewer concerned, we have expanded the evaluation to two simultaneous target clients and removal of a correctly labelled client (i.e., a benign setting).

a. Multiple target clients (Table B). On FMNIST with two label-flipping targets, PDF0.9\mathrm{PDF}_{0.9} achieves RA gaps to Retrain of 1.35 and 0.87 percentage points under α=0.1\alpha=0.1 and 1.01.0, respectively, compared with 1.31 and 0.90 points for one target. Across PDF\mathrm{PDF} variants, pooled MIA AUC differs from Retrain by at most 0.47 points. These results extend the evidence for PDF\mathrm{PDF}'s retained-utility preservation to removal of two clients.

TABLE B. Simultaneous removal results of two label-flipping target clients on FMNIST for α∈0.1,1.0\alpha\in\\{0.1,1.0\\} over four independent random seeds. Parentheses indicate signed differences from Retrain63\mathrm{Retrain}_{63}.

MetricαRetrainFUSEDCONDAE-FFedEraserFOULFedRecoveryGDFAPDF0.1\mathrm{PDF}_{0.1}PDF0.5\mathrm{PDF}_{0.5}PDF0.9\mathrm{PDF}_{0.9}
FA (%)0.10.33 ± 0.450.33 ± 0.45 (+0.00)0.75 ± 0.83 (+0.42)0.35 ± 0.43 (+0.02)0.24 ± 0.48 (−0.09)6.10 ± 7.21 (+5.77)0.48 ± 0.40 (+0.15)0.37 ± 0.42 (+0.04)0.24 ± 0.48 (−0.09)0.24 ± 0.48 (−0.09)0.48 ± 0.40 (+0.15)
1.01.45 ± 2.381.64 ± 2.32 (+0.19)1.47 ± 2.37 (+0.02)1.82 ± 2.46 (+0.37)0.72 ± 0.83 (−0.73)0.89 ± 1.18 (−0.56)1.42 ± 2.39 (−0.03)0.10 ± 0.11 (−1.35)1.40 ± 2.41 (−0.05)1.40 ± 2.41 (−0.05)0.15 ± 0.20 (−1.30)
RA (%)0.188.12 ± 1.3387.09 ± 1.74 (−1.03)85.20 ± 1.79 (−2.92)81.50 ± 1.68 (−6.62)61.62 ± 8.38 (−26.50)43.09 ± 19.89 (−45.03)85.20 ± 1.95 (−2.92)81.92 ± 2.09 (−6.20)84.18 ± 2.27 (−3.94)85.37 ± 0.78 (−2.75)86.77 ± 1.98 (−1.35)
1.090.47 ± 0.3789.93 ± 0.43 (−0.54)89.63 ± 0.06 (−0.84)86.02 ± 0.14 (−4.45)76.70 ± 1.95 (−13.77)24.77 ± 4.94 (−65.70)89.62 ± 0.17 (−0.85)88.85 ± 0.36 (−1.62)89.47 ± 0.34 (−1.00)89.57 ± 0.11 (−0.90)89.60 ± 0.45 (−0.87)
MIA AUC (%)0.150.21 ± 0.2550.25 ± 0.35 (+0.04)49.88 ± 0.34 (−0.33)50.00 ± 0.39 (−0.21)49.97 ± 0.22 (−0.24)50.08 ± 0.18 (−0.13)49.87 ± 0.28 (−0.34)50.14 ± 0.32 (−0.07)50.17 ± 0.25 (−0.04)49.89 ± 0.10 (−0.32)50.04 ± 0.19 (−0.17)
1.050.32 ± 0.3450.22 ± 0.63 (−0.10)50.06 ± 0.24 (−0.26)50.09 ± 0.20 (−0.23)50.25 ± 0.31 (−0.07)50.17 ± 0.35 (−0.15)50.00 ± 0.27 (−0.32)49.88 ± 0.46 (−0.44)49.99 ± 0.31 (−0.33)49.85 ± 0.38 (−0.47)49.91 ± 0.37 (−0.41)
Comm (GB)0.1/1.019.34 ± 0.004.11 ± 0.003.97 ± 0.007.75 ± 0.024.35 ± 0.006.08 ± 0.003.97 ± 0.003.99 ± 0.004.79 ± 0.004.79 ± 0.004.79 ± 0.00
Comp (PFLOPs)0.12.75 ± 0.021.00 ± 0.01 (−1.75)0.71 ± 0.01 (−2.04)1.21 ± 0.01 (−1.54)0.72 ± 0.01 (−2.03)0.99 ± 0.01 (−1.76)0.71 ± 0.01 (−2.04)0.71 ± 0.01 (−2.04)0.81 ± 0.01 (−1.94)0.85 ± 0.02 (−1.90)0.89 ± 0.02 (−1.86)
1.02.78 ± 0.011.01 ± 0.00 (−1.77)0.71 ± 0.00 (−2.07)1.22 ± 0.00 (−1.56)0.73 ± 0.00 (−2.05)0.99 ± 0.00 (−1.79)0.71 ± 0.00 (−2.07)0.71 ± 0.00 (−2.07)0.82 ± 0.00 (−1.96)0.83 ± 0.00 (−1.95)0.84 ± 0.00 (−1.94)

b. Benign-target removal (Table C below). On CIFAR-10 with α=0.1\alpha=0.1, PDF0.5\mathrm{PDF}_{0.5} preserves mean RA at 52.87%, compared with 52.94% before removal, showing that its utility benefit also occurs without corrupted labels. Its FA, however, falls to 22.99%, versus 59.88% for Retrain. Setting β=1\beta=1 removes target decoupling, increases RA to 59.21%, and raises FA to 50.60%, reducing the FA gap to Retrain from 36.89 to 9.28 points. This ablation demonstrates how the selection trade-off affects preservation of shared knowledge.

Target-specific MIA AUC is 50.67% for β=0.5\beta=0.5 and 52.16% for β=1\beta=1, compared with 51.82% before removal and 47.82% for Retrain. Thus, removing target decoupling improves RA and FA with Retrain, while increasing target-specific AUC by 1.49 points.

AuthorsOpenReview ↗

Response to Reviewer giuR, part 3

(Continue to respond to W2)

TABLE C. Benign-target removal results on CIFAR-10 with α=0.1\alpha=0.1 over four independent random seeds. FA is the accuracy on the target's own held-out data, whose labels are correct.

MethodParticipantsUnlearning client-roundsFA (%)RA (%)Target-only MIA AUC (%)
Original (before removal)16 of 64 per round, target included, 150 rounds–50.33±28.3350.33 \pm 28.3352.94±3.8752.94 \pm 3.8751.82±2.0051.82 \pm 2.00
Retrain16\mathrm{Retrain}_{16}16 retained per round, 150 rounds from scratch240059.88±32.3859.88 \pm 32.3846.06±12.0546.06 \pm 12.0547.82±1.3547.82 \pm 1.35
Continual FedAvg16 retained, redrawn each round, 20 rounds32032.63±16.3332.63 \pm 16.3346.17±4.4946.17 \pm 4.4950.53±3.7850.53 \pm 3.78
Fresh-random25 retained, redrawn each round, 20 rounds50044.28±25.5044.28 \pm 25.5051.35±10.8051.35 \pm 10.8052.57±5.4152.57 \pm 5.41
Fixed-random25 retained, drawn once, 20 rounds50042.46±13.7742.46 \pm 13.7747.32±3.7347.32 \pm 3.7352.02±3.5852.02 \pm 3.58
PDF0.5\mathrm{PDF}_{0.5}Top 25 by Eq. (6), 20 rounds50022.99±23.0122.99 \pm 23.0152.87±6.30\mathbf{52.87 \pm 6.30}50.67±2.9250.67 \pm 2.92
PDF1.0\mathrm{PDF}_{1.0}Top 25 by Eq. (6) with β=1\beta=1, 20 rounds50050.60±19.5350.60 \pm 19.5359.21±5.97\mathbf{59.21 \pm 5.97}52.16±6.0952.16 \pm 6.09

W3.1: Proposition 1 is a standard bound obtained by unrolling Lipschitz steps. Its weights ωt\omega_t never grow with tt, because each factor Γs\Gamma_s is at least 1. The bound does not involve FGN and does not show that timing changes the actual influence.

We agree on: Eq. (7) follows from Lipschitz unrolling, ωt=Γt+1ωt+1≥ωt+1\omega_t=\Gamma_{t+1}\omega_{t+1}\geq\omega_{t+1}, and FGN does not appear in the bound. As an upper bound, it does not establish that timing changes actual influence. However, its role is to separate the direct perturbation ∥bo(t)∥2\lVert b_o^{(t)}\rVert_2 from its propagation factor ωt\omega_t, which motivates keeping temporal information in client histories. We will revise the text after Proposition 1 to clarify this.

Regarding FGN, since ωt\omega_t alone only favors earlier rounds, any non-monotone dependence on tt must enter through bo(t)b_o^{(t)}. From Eqs. (14)–(18) on Page 15, to first order in ηℓ,t\eta_{\ell,t},

∥bo(t)∥2≈γt ηℓ,t E po(t)1−po(t) ∥go−gˉ∥2,gˉ=∑i∈S(t)pi(t)gi,\lVert b_o^{(t)}\rVert_2 \approx \gamma_t\,\eta_{\ell,t}\,E\,\frac{p_o^{(t)}}{1-p_o^{(t)}}\,\lVert g_o-\bar g\rVert_2,\quad \bar g=\sum_{i\in\mathcal S^{(t)}}p_i^{(t)}g_i,

with gradients evaluated at w−o(t−1)w_{-o}^{(t-1)}. In our one-round replays this point is the factual model w(t−1)w^{(t-1)}, at which FGN is computed (see Appendix A.2 on Page 14). There, ∑opo(t)∥go−gˉ∥22=FGN(t)/ηℓ,t−∥gˉ∥22≤FGN(t)/ηℓ,t\sum_o p_o^{(t)}\lVert g_o-\bar g\rVert_2^2=\mathrm{FGN}^{(t)}/\eta_{\ell,t}-\lVert\bar g\rVert_2^2\le\mathrm{FGN}^{(t)}/\eta_{\ell,t}. Hence a round with small FGN cannot have large direct perturbations. Specifically, the p(t)p^{(t)}-weighted mean of ∥bo(t)∥22\lVert b_o^{(t)}\rVert_2^2 is at most cγt2ηℓ,tE2FGN(t)c \gamma_t^2\eta_{\ell,t}E^2 \mathrm{FGN}^{(t)}, where c=max⁡o(po(t)/(1−po(t)))2c=\max_o\big(p_o^{(t)}/(1-p_o^{(t)})\big)^2. This is a bound and FGN is a heuristic online indicator, and the CLP claim is stated as a hypothesis (see Hypothesis 1 on Page 17).

We also test that claim by measurement. In one-round client-omission replays on FMNIST with α=0.1\alpha=0.1, the mean final parameter deviation is 1.13561.1356 for CLP rounds and 0.71760.7176 for stage-matched non-CLP rounds (paired margin 0.41790.4179, 95% CI [0.2915,0.5444][0.2915,0.5444] in Table 5 with the protocol under W3.2). The magnitude of a round-level perturbation is only weakly rank-correlated with the final deviation it causes (Spearman ρ=0.2660\rho=0.2660). Magnitude alone therefore does not determine final influence. As a result, the round in which a contribution occurs also matters, as the CLP comparison above shows.


W3.2: The support for CLP weighting therefore comes from the ablation without CLP weighting and from the FMNIST counterfactual study, and that study does not compare rounds at the same stage of training.

The appendix did not state the stage control, so we would like to clarify it here and will add it to Appendix A.4.1. In each seed, we sample up to six detected CLP rounds and pair each with its nearest non-CLP round. Both rounds in a pair use the same protocol: for three randomly chosen participants, we omit that client in that round only, replay the remaining rounds with the original participation sequence and seeds, and measure the ℓ2\ell_2 distance between the resulting and factual final models.

Across six seeds, CLP rounds exceed their matched non-CLP rounds by a paired margin of 0.41790.4179 (95% CI [0.2915,0.5444][0.2915,0.5444], tt-interval over seeds).

We agree that the other support is the non-CLP ablation, and it is setting-dependent: removing CLP weighting lowers RA by 4.884.88 points on CIFAR-10 and 0.620.62 on FMNIST at α=0.1\alpha=0.1, and by 1.331.33 and 0.090.09 at α=1.0\alpha=1.0 (see Table 3 on Page 9). CLP weighting matters most under strong heterogeneity, and we will state this in Section 5.3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author Response

Dear Area Chair and Reviewers,

Thank you for the constructive reviews and follow-up discussion. Below is a summary of our new experiments, clarifications, and manuscript revisions (highlighted in blue):

a. Isolating path-dependent selection over ordinary fine-tuning (Reviewers giuR, k8jA, dbmP). On CIFAR-10, with α=0.1\alpha=0.1, 25 retained clients for 20 unlearning rounds, PDF0.5\mathrm{PDF}_{0.5} outperforms fixed-random (+5.20% retained accuracy-RA), fresh-random (+1.75% RA), and 16-client continual FedAvg (+10.68% RA). It reduces flipped-label accuracy (FA) from 7.38% to 0.00%, compared with 4.58%, 2.93%, and 3.89% for baselines. Similar RA gains hold in the benign setting (e.g., 52.87% vs. 47.32% fixed and 51.35% fresh). We also validated simultaneous two-client removal on FMNIST, matching Retrain within 0.87-1.35% RA and 0.47% pooled MIA AUC, and clarified our stage-matched protocol for critical learning period (CLP) counterfactual replays.

b. Distinguishing client removal from label-flip correction via target-only MIA (Reviewers k8jA, dbmP). Per Eq. 4, we now evaluate FA by its absolute gap to target-excluded retraining (e.g., 0.92% to Retrain16\mathrm{Retrain}_{16} and 6.93% to Retrain63\mathrm{Retrain}_{63}) rather than treating lower FA as strictly better. We added two target-only MIAs (per-record Loss and shadow-model Shokri) on the removed client's records. Under label flipping, PDF0.5\mathrm{PDF}_{0.5} reduces target-only Loss AUC from 51.67% to 49.99%, matching Retrain16\mathrm{Retrain}_{16} (50.27%). While target-only Shokri AUC is 54.72% at β=0.5\beta=0.5, compared with 51.34% pre-removal, 49.31% Retrain16\mathrm{Retrain}_{16}, and 52.67% Retrain63\mathrm{Retrain}_{63}. Also, stronger target decoupling (β=0.1\beta=0.1) achieves 50.88%.

c. Benign client removal and shared-knowledge retention across β\beta (Reviewers dbmP, k8jA, giuR). On benign non-IID CIFAR-10, β\beta governs the trade-off between target decoupling and shared-class retention. Increasing β\beta from 0.5 to 0.9 preserves decoupling while improving target accuracy from 22.99% to 44.45% (50.60% at β=1.0\beta=1.0 compared to 59.88% for Retrain16\mathrm{Retrain}_{16}) and RA from 52.87% to 57.04% (59.21% at β=1.0\beta=1.0). Across β∈0.1,0.5,0.9,1.0\beta \in \\{0.1, 0.5, 0.9, 1.0\\}, target-only Loss AUC (47.30%–48.67%) stays within 1.07% of Retrain16\mathrm{Retrain}_{16} (48.37%), which we frame as empirical agreement with retraining given near-chance pre-removal baselines (i.e., 50.55% Loss, 51.82% Shokri).

d. Retraining baselines and communication overhead (Reviewers giuR, k8jA). We separated initial training from unlearning costs in the analysis of the manuscript. On CIFAR-10, PDF\mathrm{PDF} cuts communication by 79.2% during unlearning (39.6% total) compared with Retrain16\mathrm{Retrain}_{16}, and by 94.7% during unlearning (75.5% total) compared to full-participation Retrain63\mathrm{Retrain}_{63}.

e. Theoretical selection-utility bound (Reviewers giuR, k8jA). We clarified that Corollary 1 is a conditional worst-case guarantee valid for any ϵ≥0\epsilon \ge 0 without assuming small ϵ\epsilon a priori. Empirically, while normalized ϵaff≈0.4309\epsilon_{\mathrm{aff}} \approx 0.4309, the realized utility gap is only 8.99% of the 2kϵ2k\epsilon bound, and the PD score achieves strong client ranking (Spearman ρ=0.59\rho = 0.59), 70% overlap with the optimal subset, and 77% optimal utility recovery.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.