acceptodds
Under review as a conference paper at ICLR 2027

When Backdoors Disagree: Source-Level Federated Unlearning with Asymmetric Consensus

Abstract

Federated unlearning removes the influence of departing clients from a trained model without retraining it, and it is usually evaluated with a single deletion source. A real request may cover several groups of clients, each of which planted a different backdoor. Their deletion objectives can conflict: when the sources pull in opposing directions, one shared update may leave a source without progress, and serving all of them takes longer steps the more they cancel. We call this tension the coverage bind. A surviving source is also easy to miss, since the mean attack success over sources can be low while one backdoor persists, so we test every source separately. We propose Federated Asymmetric Consensus (FAC), whose two design choices respond to the bind: each client keeps its own objective, and the disagreement between clients and the shared model is carried across rounds rather than lost at every aggregation, and the server penalizes changes to the parameters that retained behavior relies on. The server coordinates unlearning using the forget/retain partition, without trigger specifications or source labels. In a source-scaling study with up to seven sources, FAC is the only one of sixteen approximate unlearning methods that leaves no source measurably above retraining at every source count tested. Once there are several sources, it also retains the most accuracy among the methods that do so. With three sources, FAC leaves no survivor in eight of nine residual-network runs across three datasets; on a vision transformer, neither FAC nor SalUn removes more sources on every dataset, and FAC retains more accuracy. Across 144 endpoint runs with three to seven sources, 58% show a gap of at least 20 points between worst-source and mean attack success, so averaged scores routinely hide surviving backdoors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Unlearning, Heterogeneous Backdoors, Coverage Bind, Asymmetric Consensus

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_p73AOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated unlearning, which is the process of filtering out the learned signals from poisoned clients. The deletion request when covering multiple groups of clients may result in conflicting signals. The proposed method, is an ADMM-style consensus method. The disagreement between clients and with the servers tracked and carried across rounds to be used in the unlearning process. Forget clients minimize a bounded suppression loss and retain clients minimize task loss. Per-client disagreement states are carried across rounds and also correct each client’s local proximal target. Experiments are conducted using small CNNs.ViTs on image classification tasks.

Critical strengths
  • Problem is well-motivated and the paper is well-organized.

  • Comparisons are extensive (10+ baselines) and ablations/experiment details are clear.

Critical weaknesses
  • Experiments are conducted with standard federated learning datasets such as CIFAR but more challenging tasks (or other domains such as language) could be considered.

  • Headline results and main technique illustration could be presented more cleanly. For example, based on Fig1 , the proposed technique looks heavily engineered and hard to implement. (Please don't use informal sketch style). In Table 1, only present headline setting results and leave the details to other subsections or appendix.

Questions for authors
  • The proposed technique introduces many hyper parameters. How should we tune them (considering that we may want to use the technique in OOD tasks)? Is there any aspectt of the proposed technique that is task/domain-specific or agnostic?
3 comments
AuthorsOpenReview ↗

Response to Reviewer p73A (1/2)

Thank you for the careful review. We added a text experiment and a parameter sweep, redrew Figure 1 and reorganized Table 1. Changes in the revised manuscript are marked in orange.

Response to Weakness 1: Thank you for suggesting that we extend the evaluation beyond image classification. We chose topic classification on 20 Newsgroups because we can assign each source a distinct trigger word and target topic. This lets us evaluate FAC on text using the same source-level criterion as in the image experiments. On BERT-small, FAC reused the configuration from the three-source NF-ResNet-18 experiment. None of the three sources exceeded its retraining threshold in any of the three seeds. Of the 15 baselines tested with their default settings, only SalUn also met this criterion in all three runs. FAC achieved 67.2% mean retained accuracy, 3.1 percentage points higher than SalUn’s 64.1%. The new Appendix D.3 describes the dataset and full experimental protocol.

BERT-small · 20 Newsgroups · K=3K=3

Survivor counts are listed in seed order (1/2/3). Bold marks the lowest ASR and every survivor-free entry among approximate methods.

Method  Surv ↓  Worst ASR ↓ (%)  R-Acc ↑ (%) 
No Unlearning3/3/3100.0 ± 0.066.4 ± 0.2
Retrain0/0/02.7 ± 0.666.1 ± 0.2
FT3/3/3100.0 ± 0.067.2 ± 0.2
NegGrad+3/3/3100.0 ± 0.067.2 ± 0.3
L1-Sparse3/3/3100.0 ± 0.067.2 ± 0.2
GA1/0/367.6 ± 56.166.3 ± 0.9
AggSum1/1/341.6 ± 50.865.6 ± 1.2
Sequential1/1/2100.0 ± 0.066.0 ± 0.3
PCGrad1/0/367.3 ± 56.666.3 ± 0.8
MGDA0/2/339.5 ± 52.866.1 ± 0.5
FedOSD1/0/367.6 ± 56.167.0 ± 0.3
FedCARE1/0/367.5 ± 56.266.8 ± 0.1
RFE1/1/368.3 ± 54.967.1 ± 0.3
FedRazor1/0/367.3 ± 56.766.1 ± 1.2
GDFA-U3/3/3100.0 ± 0.066.7 ± 0.1
NoT2/1/378.0 ± 18.363.3 ± 1.4
SalUn0/0/02.8 ± 0.764.1 ± 0.4
FAC0/0/00.8 ± 0.967.2 ± 0.5

Response to Weakness 2: Figure 1 now shows one federated round as a vector diagram, and Table 1 now keeps only the results for three, five and seven sources. The figure follows the server broadcast, local training and server aggregation. On the client side, it shows where local correction occurs. On the server side, it shows how the previous round’s state enters the current aggregation and how the current round’s difference carries forward. Appendix A gives the formulas and pseudocode. The single-source results are now in Appendix C. Table 1 retains all 18 methods.

AuthorsOpenReview ↗

Response to Reviewer p73A (2/2)

Response to Question 1: For a new task, we recommend starting from the closest tested configuration and prioritizing the regularization strength λ\lambda and the initial unlearning rate η0\eta_0. The table shows the sweep in our main setting (non-IID CIFAR-10, CNN, five sources, three seeds). Appendix E gives tuning guidance and the full sweep.

  • Settings that met the deletion criterion throughout the sweep. Every tested value of α\alpha, ωf\omega_{\mathrm{f}} and δ\delta, and every positive value tested for ρp\rho_{\mathrm{p}}, left no surviving sources in all three seeds. Retained accuracy and worst-source ASR still varied across these settings. With ρp=0\rho_{\mathrm{p}}=0, sources crossed back above threshold after recovery.
  • Parameters that need no separate tuning. The server update depends on ρ\rho and λ\lambda only through their ratio λ/ρ\lambda/\rho. Doubling both reproduced the original results exactly, so it suffices to tune λ\lambda.
  • Parameters to prioritize on a new task. Start with λ\lambda and η0\eta_0. On the CNN, small values of λ\lambda left surviving sources (the rows with λ≤10\lambda\le10). We recommend varying one parameter at a time and using held-out data to select a setting with lower worst-source ASR among those that meet the retained-accuracy requirement. BERT-small reused the three-source NF-ResNet-18 values without further adjustment.

The server computations are task-independent. We use the same across-round disagreement, consensus and sensitivity calculations for images and text. The task-specific choices on the client side are the trigger applied to inputs (an image pattern or a text token), the target class and the unlearning loss, which is designed for classification. To adapt the computation to another type of task, only the clients’ unlearning loss needs to be replaced.

Parameter sweep · CNN · CIFAR-10 · non-IID · K=5K=5 · 50 unlearning + 50 recovery rounds

Variants change one parameter from the first-row configuration, except for the final row, which doubles both ρ\rho and λ\lambda to keep their ratio fixed. Rows marked * are new runs; the others come from the paper's Appendix E. The remaining sweep values are in Appendix E.

Parameter  Value  Surv ↓  Worst ASR ↓ (%)  R-Acc ↑ (%) 
FACλ=100,ωf=3,α=0.05,δ=0.1,ρp=0.05,ρ=10\lambda=100,\\ \omega_{\mathrm{f}}=3,\\ \alpha=0.05,\\ \delta=0.1,\\ \rho_{\mathrm{p}}=0.05,\\ \rho=100.001.98±1.3850.83±0.32
λ\lambda12.6767.84±16.5447.57±2.81
λ\lambda101.3385.05±19.3546.27±1.45
λ\lambda500.000.99±0.3550.30±0.79
λ\lambda2000.002.52±1.4250.80±0.36
ωf\omega_{\mathrm{f}}10.002.82±1.3951.00±0.46
ωf\omega_{\mathrm{f}}100.001.22±0.7049.63±0.49
α\alpha0.010.005.18±3.3351.93±0.49
α\alpha0.500.002.21±1.2648.73±0.90
δ\delta *0.010.002.36±1.6250.33±0.31
δ\delta *10.001.22±0.4750.57±0.31
ρp\rho_{\mathrm{p}}01.3315.32±1.6552.93±0.35
ρp\rho_{\mathrm{p}} *0.010.005.34±1.5152.90±0.44
ρ,λ\rho,\\ \lambda *ρ=20,λ=200\rho=20,\\ \lambda=2000.001.98±1.3850.83±0.32

I thank the authors for the detailed rebuttal. They address my concerns and I raise the rating to 3.

Reviewer_p7wVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This work reveals a blind spot in evaluating federated unlearning when facing multiple distinct backdoors. Even after unlearning decreases the average attack success rate to the same level as that of the retrained reference model, the unlearned model can still be exposed to strong risks from one specific backdoor. This work proposes Federated Asymmetric Consensus (FAC), a mechanism that leverages persistent client–consensus disagreement states with regularization to adjust the updates. The theoretical analysis shows the difficulty of formulating such objectives. Experiments are run across different datasets and architectures. The ablation study shows the effectiveness of each component

Critical strengths
  1. The problem is critical and practical to explore. The underlying blind spot is valid when different poisons are encountered.

  2. The trade-off is promising for the defined settings.

  3. The empirical analysis is extensive and shows a carefully curated comparison of the components. It clearly shows that the assignment of sensitivity scores to coordinates matters while their distributions and selected-coordinate counts are preserved

Critical weaknesses
  1. The comparative advantage is confounded by the configuration choice. Table A6 provides a useful partial control, but it also demonstrates that configuration choices materially affect the conclusions. A more balanced tuning budget for the strongest baselines, particularly on the larger backbones, would make the comparative evidence more convincing.

2.Operational justification is needed to embed the practical setting. FAC starts with a supplied forget-client set and requires the departing clients to cooperate in local unlearning using their own trigger–target pairs. Nevertheless, there should be a discussion of why it is a practical scenario to address in a real-world workflow.

  1. The paper is a bit difficult to follow. The jump from the average attack success rate to gradient geometry and the consensus mechanism occurs without sufficient bridging of the content.
Questions for authors
  1. Is there any concrete evidence to support this mechanism, and is there a real-world workflow that motivates it? For example, when and how will FAC be triggered?

  2. Can the strongest baselines be compared under matched tuning budgets, including method-specific parameters and recovery choices? How much of FAC’s advantage remains under such a comparison, particularly on the larger backbones?

  3. How should the successful merged-state and no-local-correction configurations affect the interpretation of the proposed mechanism?.

3 comments
AuthorsOpenReview ↗

Response to Reviewer p7wV (1/2)

Thank you for the careful review. We added tuning comparisons with matched per-run budgets on both larger backbones, described the practical workflow, and revised our explanation of the ablations. Changes in the revised manuscript are marked in orange.

Response to Weakness 1: On both larger backbones, we tested four strong baselines (SalUn, NoT, FedOSD and FedCARE) with 100 rounds per run, varying one method-specific setting or the stage allocation at a time. FAC retained higher accuracy at equal deletion on NF-ResNet-18. On ViT, default SalUn had the highest retained accuracy among survivor-free baselines (Q2).

Response to Weakness 2: FAC addresses institutional withdrawal when the partners have agreed on a handover. A data provider requests withdrawal of its training contribution and agrees with the project team to carry out limited local cleanup before disconnecting its node. The project agreement template that Medical Data Works publishes for its healthcare federation platform, Railway, calls on partners to provide reasonable transition assistance for ongoing projects when a member withdraws [1, §14.1(iii)]. Van Soest et al. report that 11 project or consortium agreements have been signed under this governance framework [2].

In our backdoor setting, the operator responsible for an affected client seeks to withdraw its contribution, can still run computations locally, and has identified the trigger and its target class. Participating in unlearning helps the provider carry out the withdrawal and correct the problem it found. The partners can agree to do this work during the handover. FAC coordinates updates from the departing and retained clients during this phase. Only retained clients participate in any subsequent training.

Response to Weakness 3: We expanded the explanation of why source-level evaluation leads us to analyze gradient directions. Even when mean ASR falls, testing each source can reveal a backdoor that still works. Even if the corresponding client takes another local step, updates from other sources may point in opposing directions and partly cancel its update during aggregation. Three passages in the revised manuscript explain how this observation motivates the gradient analysis and how that analysis guides FAC’s consensus update.

  1. The introduction explains why each source is evaluated separately. If four of five sources have 0% ASR and the fifth has 100%, mean ASR is only 20%, yet the fifth backdoor works every time.
  2. Section 3.1 explains why update directions matter before introducing Proposition 1. The server broadcasts the same shared model each round. One client may need to increase a parameter while another needs to decrease it. Both can make progress locally, yet aggregation can return the parameter almost to its starting value. Proposition 1 characterizes when the updates cancel.
  3. Sections 3.2 and 3.3 explain what each design addresses before giving its equation. Before Eq. (4), we explain how a change partly cancelled in one round can still influence the next. Before Eq. (8), we explain why some parameters can change more while those supporting the retained task should stay close to their pre-unlearning values.

Response to Question 1: FAC addresses multi-source unlearning during the withdrawal transition described in W2. Affected clients or client groups carry distinct backdoors with locally identified trigger–target pairs, and their legitimate operators cooperate in cleanup before disconnecting. FedOSD, SFU and FedCARE also involve forget clients in local computation [3–5].

  • When is FAC triggered? FAC starts once the server receives the withdrawal requests and determines the forget-client set.

  • How is FAC triggered? The server broadcasts the current shared model. Forget clients perform local unlearning updates, while retain clients perform normal task updates. FAC carries client–consensus disagreement across rounds to coordinate potentially conflicting updates from multiple sources. After unlearning, only retain clients participate in subsequent recovery or training.

[1] Medical Data Works. Federated Learning Agreement Template for [CONSORTIUM] projects. Version 20260207, §§14.1–14.2, p. 22.

[2] Johan van Soest, Luis Sanchez Gomez, Leonard Wee, and Andre Dekker. A Governance Framework for Federated Learning Projects in Healthcare. Medical Informatics Europe (MIE 2025), Studies in Health Technology and Informatics, vol. 327, pp. 258–262. DOI 10.3233/SHTI250324.

[3] Zibin Pan et al. Federated Unlearning with Gradient Descent and Conflict Mitigation. AAAI 2025, (FedOSD).

[4] Guanghao Li et al. Subspace based Federated Unlearning. Transactions on Machine Learning Research, 2025 (SFU).

[5] Yue Li et al. FedCARE: Federated Unlearning with Conflict-Aware Projection and Relearning-Resistant Recovery. IJCAI 2026.

AuthorsOpenReview ↗

Response to Reviewer p7wV (2/2)

Response to Question 2: On NF-ResNet-18, FAC retained higher accuracy at equal deletion. On ViT-S/16 with CIFAR-10, default SalUn had zero survivors in all three seeds at 95.10% retained accuracy. FAC retained 95.47% with one survivor in one seed. The additional settings preserve the trade-off reported in the submission.

All runs used the same three CIFAR-10 sources, three seeds and 100 rounds. FAC kept its submitted settings. Each baseline variant changed one setting. We varied SalUn’s mask ratio, the scope of NoT’s weight negation, FedCARE’s recovery-filter strength, and the split between unlearning and recovery for SalUn, FedOSD and FedCARE. The number of settings tested varied by method.

SalUn illustrates why this comparison matters. On NF-ResNet-18, its default schedule of 50 unlearning rounds and 50 recovery rounds left one surviving source in every seed. Using all 100 rounds for unlearning left zero survivors in all three seeds, but retained accuracy fell by about 19 percentage points. We therefore compare retained accuracy among configurations with zero survivors in all three seeds. The best baseline under this criterion was NoT at 70.00%, compared with 76.59% for FAC. The tables below list every configuration, as in Appendix D.2.

Bold marks the highest mean R-Acc among settings with 0/0/0 survivors.

NF-ResNet-18 · CIFAR-10 · K=3K=3

Method Setting  Surv ↓  Worst ASR ↓ (%)  R-Acc ↑ (%) 
FACsubmitted0/0/01.12±0.1076.59±0.06
SalUnratio 0.5; 50+5050+501/1/16.10±2.7650.42±1.23
SalUnratio 0.1; 50+5050+502/2/223.03±9.3751.73±0.75
SalUnratio 0.5; 100+0100+00/0/00.93±0.1931.13±1.27
NoTall tensors0/0/02.94±1.5670.00±0.21
NoTfirst tensor only2/2/117.46±6.8475.40±0.61
FedOSD50+5050+501/2/193.85±7.7376.09±0.26
FedOSD100+0100+01/2/186.35±15.6774.83±0.74
FedCARE50+5050+501/2/194.80±5.4875.37±0.36
FedCAREhalf filter; 50+5050+501/2/195.09±5.1275.45±0.31
FedCARE100+0100+01/2/051.88±45.2072.19±1.78

ViT-S/16 · CIFAR-10 · K=3K=3

Method Setting  Surv ↓  Worst ASR ↓ (%)  R-Acc ↑ (%) 
FACsubmitted1/0/03.20±5.4295.47±0.47
SalUnratio 0.5; 50+5050+500/0/00.30±0.3595.10±0.69
SalUnratio 0.1; 50+5050+500/0/00.48±0.3694.67±0.45
SalUnratio 0.5; 100+0100+00/0/00.00±0.0083.67±0.60
NoTall 152 tensors0/0/00.00±0.0011.93±0.81†
NoTfirst tensor only3/3/390.50±2.1595.20±0.50
FedOSD50+5050+502/2/217.11±7.4695.70±0.35
FedOSD100+0100+00/1/14.79±1.2595.70±0.36
FedCARE50+5050+502/2/217.63±8.6395.70±0.35
FedCAREhalf filter; 50+5050+502/2/217.69±8.7295.70±0.35
FedCARE100+0100+00/2/25.26±1.1395.70±0.44

Here, ratio denotes SalUn’s mask ratio, half filter halves FedCARE’s recovery-filter strength, a+ba+b denotes aa unlearning and bb recovery rounds, and † marks R-Acc below 30% in every seed.

Response to Question 3: Forget clients can share one state, and local correction matters after recovery. The state uku_k is the running sum, kept on the server, of client kk's returned model minus that round’s consensus. In full FAC the proximal target is z−ukz-u_k.

 Merged state. With one state shared by all forget clients (target z−uFz-u_{\mathcal F}), zero sources survived at round 100, as in full FAC. Removing the state left 2.33 survivors on average at both round 50 and round 100, so what matters is having the state, not storing it per client. The target matters too: with the same shared state, using zz as the proximal target left 1.33 survivors.

 Local correction. Both variants kept individual server-side states. Disabling the local proximal step (ρp=0\rho_{\mathrm p}=0) left zero survivors at round 50 and 1.33 on average after recovery. Keeping the step but changing its target from z−ukz-u_k to zz left zero at round 50 and 1.00 after recovery. Table A18 in Appendix E gives both endpoints.

These results are for the CIFAR-10 CNN at K=5K=5 with 50 unlearning rounds followed by 50 recovery rounds.

Configuration  R50 Surv ↓  R100 Surv ↓  R100 R-Acc ↑ (%) 
Full FAC0.000.0050.83
Merged forget state, target z−uFz-u_{\mathcal F}0.000.0050.77
Merged forget state, shared target zz0.001.3352.90
Individual client states, shared target zz0.001.0052.93
No local correction (ρp=0\rho_{\mathrm p}=0)0.001.3352.93
No across-round state (u=0u=0)2.332.3353.23

Dear authors, Thank you for your response. I believe the results address most of my concerns and strengthen the paper. I will keep my overall score unchanged and raise my soundness score.

Thank you again for your efforts.

Reviewer_21YWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a federated unlearning algorithm called FAC to resolve conflicting gradients from unlearning clients. For each client, FAC stores the accumulated difference (or "disagreement") between its local unlearned model and the global model across unlearning rounds. The server in FAC uses the ADMM consensus framework to resolve these disagreements during aggregation. Additionally, FAC penalizes the ADMM-style aggregation objective function based on the difference between parametric offsets induced by the forget clients and those induced by the retain clients.

Critical strengths
  1. The problem of conflicting unlearning gradients is well-motivated. The experimental results corroborate the ineffectiveness of existing methods in reconciling conflicting gradients from unlearning clients.

  2. FAC shows strong empirical performance in unlearning multi-source triggers by resolving the conflicting gradients. The empirical advantages hold consistently across different data partitions, models, and datasets.

Critical weaknesses
  1. The paper does not discuss the additional cost (e.g., communication overhead, computation cost and storage overhead at the server) of FAC.

  2. The writing, particularly in Sections 3.2 and 3.3, can be improved by providing motivation and justification for each step, such as Eq. 4 and Eq. 8. This could make the paper more readable for those who are not familiar with the ADMM consensus algorithm.

Questions for authors
  1. Can the authors elaborate on the additional cost of FAC in terms of communication overhead, as well as the computation cost and memory overhead at the server?
  2. FAC carries and resolves gradient disagreement across multiple unlearning rounds. How many iterations does FAC often need in practice? Can the authors show the performance changes when varying the number of unlearning rounds?
  3. (Clarifying question) In Table 1, why does the retrained model not achieve the lowest attack success rates?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 21YW (1/2)

Thank you for the careful review. We added analytical and measured costs, experiments with shorter unlearning schedules, and a paired analysis of Retrain’s predictions. Changes in the revised manuscript are marked in orange.

Response to Weakness 1: Our implementation adds two costs. The server stores one model-sized state per client, the running sum of the differences between that client’s returned model and the consensus model of each round. With local correction enabled, each client also downloads one extra model-sized target per round. Server updates were fast, taking 5.5 seconds in total over a 100-round ViT run. With full participation and fixed client weights, the clients can store their own states and avoid the extra download, as described in Q1. Appendix B.3 of the revised manuscript gives the details.

Response to Weakness 2: Sections 3.2 and 3.3 of the revised manuscript explain the problem each step addresses before introducing the corresponding equation.

Section 3.2 explains why FAC carries disagreement across rounds. Updates from different sources can partly cancel during aggregation, so the consensus model may reflect only part of a client’s update. To carry the part of a client’s update that was not reflected in the consensus into later rounds, FAC accumulates the difference between each client’s local model and the consensus model at each round. Equation (4) incorporates these accumulated differences into the next consensus update. With local correction enabled, client kk also uses its state uku_k to form its local target z−ukz-u_k.

Section 3.3 explains how the consensus update balances deletion and retained accuracy. Some parameters can change more than others, while those on which the retained task relies should stay close to their pre-unlearning values. The server compares the magnitudes of forget- and retain-client updates at each coordinate. Equation (8) then shrinks selected high-scoring coordinates toward zero and anchors low-scoring coordinates outside the selected region more strongly to their pre-unlearning values.

Response to Question 1: For a model with PP parameters of bb bytes each and NN clients, the costs are as follows, in the order you asked about them. Appendix B.3 of the revised manuscript includes these figures.

  • Communication. Each round, a FedAvg client downloads the shared model and uploads its local model. With local correction enabled, a FAC client also downloads its correction target. This gives three model-sized transfers instead of two, increasing the payload for those rounds by 50%. We enable local correction during unlearning on the CNN and NF-ResNet-18. Our ViT runs disable it and have the same communication cost as FedAvg.
  • Server computation. Each round, the server updates the client states and the consensus model coordinate by coordinate. This takes O(NP)\mathcal{O}(NP) work, the same order as FedAvg’s weighted average. The table below gives the measured times.
  • Server memory. Our implementation stores each client’s state on the server. Each state accumulates the difference between the returned and consensus models across rounds. Together these states take bNPbNP bytes, about 0.87 GB for ViT-S/16 with ten clients and 32-bit floats.

To reduce server memory use or avoid the extra download, the clients can store their own states. With full participation in every round and fixed client weights, each client can compute its own correction target, and the server need only maintain the weighted sum of the states. Server storage for these states then falls to one model-sized vector (bPbP bytes), and communication matches FedAvg in all configurations.

 Timing on ViT-S/16 · CIFAR-10 · 100 rounds

Method  Client total  Parallel clients (est.)  Server updates 
FAC2,4703395.5
FedOSD2,15933528.9
SalUn2,12133321.3
NoT2,12633929.5
AuthorsOpenReview ↗

Response to Reviewer 21YW (2/2)

Response to Question 2: In our main non-IID CIFAR-10 experiment, five unlearning rounds sufficed for the CNN to meet the source-wise deletion criterion at both K=5K=5 and K=7K=7. This was the shortest schedule we tested. At each source count, all three seeds had zero surviving sources at the end of unlearning and again after 50 recovery rounds.

With recovery fixed at 50 rounds, all four unlearning schedules (5, 10, 20 and 50 rounds) met this criterion. Compared with the original 50+5050+50 schedule, 10+5010+50 reduced total communication from 100 to 60 rounds. Its mean worst-source ASR after recovery was slightly higher (2.29% vs. 1.98% at K=5K=5 and 1.83% vs. 1.60% at K=7K=7). Retained accuracy was similar (50.0% vs. 50.8% at K=5K=5, and 47.2% for both at K=7K=7). Appendix G (Table A24) of the revised manuscript reports these runs.

 K  Unlearning rounds  Survivors  Worst ASR (%)  R-Acc (%) 
550/0/02.90 ± 1.9449.8 ± 0.6
5100/0/02.29 ± 1.5050.0 ± 0.8
5200/0/02.44 ± 1.8350.1 ± 0.4
5500/0/01.98 ± 1.3850.8 ± 0.3
750/0/04.18 ± 2.0246.6 ± 0.5
7100/0/01.83 ± 0.6247.2 ± 0.2
7200/0/01.70 ± 0.4547.1 ± 0.2
7500/0/01.60 ± 0.4247.2 ± 0.5

Survivors are listed in seed order (1/2/3). Worst-source ASR and retained accuracy are measured after recovery.

Response to Question 3: Most of the predictions counted toward Retrain’s ASR already occur without the trigger. Retrain never trains on data from the backdoored clients, but like any classifier it sometimes assigns a non-target image to the target class. If such an image is still classified as the target class after the trigger is added, it counts toward ASR.

We ran the same Retrain model twice on each non-target image, first on the original and then with the trigger added. In the five-source setting, Retrain made 229 target-class predictions on triggered images. Of these, 174 were already target-class predictions on the original images. Only 55 switched to the target class after the trigger was added, so roughly three quarters would have occurred without it.

An unlearning method that actively suppresses target-class predictions can fall below this level. This explains why some methods in Table 1, including FAC, have lower ASR than Retrain. It also explains why we use Retrain to calibrate each source’s survival threshold instead of requiring zero ASR. We have added this paired evaluation to Appendix B.3 of the revised manuscript.

 K  Clean (%)  Triggered (%)  Into target  Out of target 
13.64.6144
34.33.72661
54.03.555108
75.45.6166184

Clean and Triggered report target-class prediction rates, averaged first across clients within each source, then across sources and seeds. Into target and Out of target count all probes whose prediction switched into or out of the target class when the trigger was added.

Thank you for your detailed responses. I have the remaining concerns:

[W1 + Q1] Regarding the memory cost, the authors mentioned that the server needs to store one additional model state for each client. Since memory scales linearly with the number of clients, this could quickly get expensive when the number of clients is large (>= 100) and the base model has a large number of parameters (e.g., LLMs). I suggest the authors should be explicit about this limitation in their paper.

Regarding the communication cost, it is misleading to say your method has the same communication cost as FedAvg as long as the local correction step is disabled. Additionally, why do you disable it in your ViT run? Similar to the memory cost, I suggest that the authors explicitly mention their additional communication cost in the main paper, rather than in the appendix section.

Response to Question [W1+Q1]: Thank you for following up on these points.

Server memory. We agree that this is a limitation. Our current implementation stores one parameter-sized state per client on the server, adding bNPbNP bytes, which grows linearly with both the number of clients and the model size. With 100 clients and a one-billion-parameter model in 32-bit floats, these states alone would take 400 GB. Section 3.2 of the revised manuscript now states this storage cost, and Section 5 lists it as a limitation for large federations and large models.

Communication. We agree that saying FAC has the same communication cost as FedAvg when local correction is disabled was misleading. In our current implementation, which keeps the states on the server, local correction (ρp>0\rho_{\mathrm p}>0) requires the server to send each participating client an additional correction target. In those rounds, each client transfers 3bP3bP instead of FedAvg’s 2bP2bP, a 50% increase, and its download doubles from bPbP to 2bP2bP. Section 3.2 of the revised manuscript now states this additional cost in the main text.

Why the ViT runs disable local correction. Like the other coefficients in Table A1, ρp\rho_{\mathrm p} was set per protocol during development (Appendix B.1). On ViT-S/16, the largest image backbone in the paper, the main table uses the lower-communication setting ρp=0\rho_{\mathrm p}=0, which transfers 2bP2bP per client per round. To quantify this trade-off, we repeated all nine ViT runs with ρp=0.05\rho_{\mathrm p}=0.05, the value used on CNN and NF-ResNet-18, under the same protocol, three seeds and 100-round budget. With correction enabled, all nine runs were survivor-free, the original configuration’s only surviving source on CIFAR-10 fell below its threshold, and mean retained accuracy matched or slightly exceeded the original on all three datasets. On ViT, local correction therefore buys further deletion at the cost of one extra model-sized download per client in each correction round. The main table keeps the submitted configuration for consistency with the original results, and Table A3(b) in Appendix B.3 reports the comparison.

 Dataset  ρp\rho_{\mathrm p}  Surv (s1/s2/s3)  Worst-source ASR (%)  R-Acc (%) 
CIFAR-1001/0/03.20 ± 5.4295.47 ± 0.47
CIFAR-100.050/0/00.95 ± 1.5295.73 ± 0.67
CIFAR-10000/0/00.35 ± 0.4683.07 ± 0.15
CIFAR-1000.050/0/00.31 ± 0.2183.30 ± 0.20
SVHN00/0/00.94 ± 1.4893.60 ± 0.10
SVHN0.050/0/00.00 ± 0.0093.63 ± 0.12

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.