acceptodds
Under review as a conference paper at ICLR 2027

When the Bubble Bursts: Malicious Unlearning Requests Undermine Federated Learning

Abstract

Growing demand for the “right to be forgotten” has motivated federated unlearning (FU), which aims to efficiently remove the influence of designated training data while preserving model utility. However, we identify a security gap in existing FU approaches: treating client-issued unlearning requests as trusted inputs allows adversaries to exploit the unlearning service itself to manipulate model predictions. To expose this vulnerability, we propose BURST, an attack framework that induces misclassification of selected samples through malicious unlearning requests. Specifically, the adversary constructs a target-support set and an interference set from public auxiliary data and incorporates both into federated training, using these external samples to exert controllable adversarial influence on the target prediction. The support set reinforces the model’s correct prediction for the target sample, while the interference set is designed to undermine it. During the unlearning phase, the adversary submits malicious requests to unlearn the support set and strengthens the interference set’s adversarial influence, jointly turning previously correct predictions into misclassifications. Extensive experiments show that BURST achieves an attack success rate of 91.43% on CIFAR-10 while reducing global model accuracy by only 5.66%. Among misclassified targets, confidence in the predicted labels is on average 30.74% higher than confidence in the ground-truth labels. Our findings demonstrate how client-controlled unlearning requests and data can jointly compromise FU, highlighting the need to validate both to safeguard model integrity. Code is available at https://anonymous.4open.science/r/Code-for-BURST-4CC1.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Federated Unlearning, Malicious Unlearning Requests

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5x8VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates security vulnerabilities in federated unlearning (FU), focusing on how malicious clients can manipulate model predictions through coordinated training-data modifications and unlearning requests. The authors propose BURST, an attack framework that constructs two complementary sets from public auxiliary data: a target-support set that reinforces correct predictions and an interference set that promotes competing-class predictions. During federated training, both sets are incorporated into malicious clients’ local datasets. During unlearning, the attacker removes the support set and strengthens the interference set through optimized perturbations, aiming to turn previously correct predictions into misclassifications. Experiments on MNIST and CIFAR-10 demonstrate the effectiveness of BURST against existing attacks, including BadFU and FedMUA. On CIFAR-10, BURST achieves a 91.43% attack success rate while reducing global accuracy by 5.66 percentage points.

Critical strengths

S1. The paper addresses an important vulnerability in federated unlearning, highlighting that unlearning requests and subsequent model reconstruction can introduce security risks beyond conventional training-time attacks.

S2. The combination of target-support removal and interference enhancement provides a coherent mechanism for manipulating predictions specifically during the unlearning stage.

S3. BURST achieves substantially higher attack success rates than the evaluated baselines, particularly on CIFAR-10 (91.43% compared with 60.00% for FedMUA).

S4. BURST avoids the expensive Hessian-based influence estimation required by FedMUA and demonstrates substantially lower attack-construction overhead.

Critical weaknesses

W1. The attack assumes access to the target sample and its ground-truth label, a suitable public auxiliary dataset, and the ability to manipulate client-side reconstruction or recalibration. These assumptions may limit its applicability to practical FU systems.

W2. Although the paper identifies security vulnerabilities, it does not experimentally investigate whether existing malicious-update detection, data validation, or robust aggregation mechanisms can mitigate BURST.

W3. The reported runtime comparison excludes surrogate-model training, federated training, and unlearning. Consequently, the results do not establish the end-to-end computational advantage of BURST.

Questions for authors

Q1. Section 3.1 defines an unlearning request as r=(F,q), where F is the forget set and q contains protocol-specific information. However, Algorithm 1 constructs the malicious request as \tilde r=(S,\tilde I). It is unclear how the enhanced interference set is incorporated into the standard FU interface and whether the assumed FU methods permit this operation. This distinction is important because the attack relies on manipulating both removal and reconstruction.

Q2. The methodology constructs attack sets before federated training, whereas Section C.4 describes selecting target samples based on their correct classification by the pre-unlearning model. The paper should clarify whether targets are selected using a separately trained reference model or whether attack construction and target filtering are performed iteratively. This matters because target selection could introduce evaluation bias.

Q3. The paper requires p_i\leq k_i, but this does not guarantee that the selected competing class contains at least p_i samples within the retrieved neighborhood. The authors should clarify how the algorithm handles cases where fewer than p_i samples belong to the selected class.

Q4. The analysis defines the target margin relative to the strongest competing class before unlearning. However, this competing class may change after unlearning. Consequently, the statement that a nonpositive margin necessarily characterizes post-unlearning misclassification is valid as a sufficient condition, but not generally as an equivalent characterization. The distinction should be clarified.

Q5. The paper states that attack-specific budgets are matched across methods, but Appendix C.3 qualifies this as being done “whenever possible.” The authors should explicitly report the actual sample budgets for each baseline to establish the fairness of the comparisons.

Q6. Table 2 reports only 1.041 seconds for BURST’s attack construction, while the methodology requires surrogate-model training and feature-space retrieval. It is unclear whether these operations are included in the reported runtime. Additionally, Appendix A models perturbation optimization using a second-order operation cost, although the proposed optimization does not explicitly require second-order derivatives. These aspects should be clarified.

Q7. FedMUA achieves a substantially higher confidence gap than BURST on CIFAR-10 (86.38 versus 30.74 percentage points), despite having a lower attack success rate. The paper emphasizes BURST’s effectiveness but does not sufficiently discuss this trade-off between attack success frequency and misclassification confidence.

Code README.md should be written English rather than Chinese.

10 comments
AuthorsOpenReview ↗

Response to Reviewer 5x8V (1/5): W1 & W2 - Threat Model and Defenses

Critical Weakness 1: Practical Applicability of the Threat Model

Access to the target sample and its ground-truth label. These assumptions correspond to targeted attack scenarios in which routine collaboration gives two participating clients access to the same labeled target. In hospital referrals, an authorized transfer of a dermoscopic image and its pathology-confirmed diagnosis allows both the referring and receiving hospitals to hold the same target and label. An attacker controlling participating nodes aims to compromise the shared model’s automated assessment of that case.

In collaborative phishing detection, security organizations can share the same webpage screenshot and its verified phishing label through joint investigations and threat-intelligence exchange. MISP supports such cross-organizational information sharing and image attachments. (MISP information sharing; MISP image objects) Here, the attacker aims to make the post-unlearning classifier misclassify a specific phishing page as legitimate, undermining downstream automated detection. In both scenarios, knowing the target’s label during attack construction does not mean that the model receives it during inference.

Public auxiliary-data sources. Publicly available ISIC2019 dermoscopic images carry diagnostic labels and provide a potential auxiliary-data source for skin-lesion classification; FLamby’s Fed-ISIC2019 provides a research benchmark for this federated task. (Fed-ISIC2019) For phishing detection, LNU-Phish provides screenshots and class information for legitimate and phishing webpages. (LNU-Phish) When the input formats and class definitions are compatible, these resources provide candidate samples for constructing the support and interference sets without accessing other participants’ private training data.

Control over client-side calibration. In these scenarios, we consider federated unlearning systems that support sample-level deletion and allow the relevant clients to continue participating in local calibration. Malicious clients incorporate support and interference samples during training, subsequently request deletion of the support records, and compute local calibration updates using enhanced interference samples with unchanged labels. The server receives model updates and performs the established calibration and aggregation procedures. These operations use malicious clients’ control over their own data and deletion requests, requiring neither server control nor additional server interfaces.


Critical Weakness 2: Evaluation against Existing Defenses

We evaluate BURST against Krum, Trimmed Mean, and FLAME on MNIST, distinguishing defenses applied during federated training from those applied during unlearning. BURST retains ASRs of 60.00%–75.00% under the evaluated training-stage defenses and 68.18%–72.73% under the evaluated unlearning-stage defenses.

Defense settings. We evaluate Krum, Trimmed Mean, and FLAME when applied only during federated training. We separately evaluate Trimmed Mean and FLAME when applied only during unlearning, directly examining whether defenses at the reconstruction and calibration stage prevent the attack. The results are as follows:

Defense stageDefenseASR (%)Final test accuracy (%)
Training onlyKrum75.0091.39
Training onlyTrimmed Mean63.6496.01
Training onlyFLAME60.0095.70
Unlearning onlyTrimmed Mean72.7395.45
Unlearning onlyFLAME68.1895.37

Attack effectiveness under defenses. When defenses operate only during training, BURST achieves ASRs of 75.00%, 63.64%, and 60.00% against Krum, Trimmed Mean, and FLAME, respectively. When defenses operate only during unlearning, ASRs are 72.73% against Trimmed Mean and 68.18% against FLAME. Thus, successful attacks persist even when defenses are applied directly at the unlearning stage.

The evaluated defenses do not eliminate BURST in these settings, motivating protection that explicitly accounts for attacks during unlearning. In the revised manuscript, we report these defense evaluations and distinguish their placement during training and unlearning.


AuthorsOpenReview ↗

Response to Reviewer 5x8V (2/5): W3 & Q1 - Runtime Scope and FU Interface

Critical Weakness 3: Scope of the Runtime Comparison

Table 2 reports attack-construction time with a prepared surrogate model. The efficiency comparison therefore concerns the attack-construction stage, excluding surrogate-model training, federated training, and unlearning execution.

Surrogate preparation and reuse. The surrogate model is reusable across targets within the same task and does not require retraining for each target. A compatible pretrained model can also serve as the surrogate. We therefore distinguish surrogate preparation from per-target attack construction when reporting computational costs.

Shared training and unlearning protocols. All compared attack methods use the same federated training and unlearning protocols and configurations. Neither BURST nor the compared baselines modifies the underlying federated training or unlearning protocol; their attack-specific operations concern sample construction and the use of these samples within the prescribed workflow. Table 2 focuses on the computational overhead of attack construction within this shared experimental framework.

In the revised manuscript, we explicitly define the timing scope in Section 5.4 and Appendix C.6.


Q1: Unlearning Requests and Client-Side Interference Enhancement

The enhanced interference set I~\tilde I is not part of the unlearning request submitted to the server. The request is r=(S,q)r=(S,q), while interference enhancement is performed locally by the interference-holding client during calibration. These are separate operations within the stated threat model.

In the revised manuscript, we align Algorithm 1 with the interface defined in Section 3.1 as follows:

1. Request representation. We retain the definition r=(F,q)r=(F,q) in Section 3.1 and replace r~=(S,I~)\tilde r=(S,\tilde I) in the request-construction step of Algorithm 1 with r=(S,q)r=(S,q). Here, F=SF=S, and qq contains only protocol-required information. The enhanced interference set I~\tilde I is not included in the request.

2. Algorithm execution. We expand the original lines 10–11 into the following steps:

  1. Submit the unlearning request r=(S,q)r=(S,q).
  2. When the FU procedure invokes local calibration, the client holding the support set uses its remaining data after removing SS. The client holding the interference set locally replaces II with its label-preserving enhanced version I~\tilde I and computes updates through the prescribed local training procedure.
  3. The server performs calibration and aggregation based on the request and the returned client updates, producing w~−=U(w~T,r)\tilde w^{-}=\mathcal U(\tilde w_T,r).

3. Corresponding descriptions in the main text and appendix. In Section 4.3 and Appendix B, we describe the two operations as follows:

The unlearning request identifies SS as the forget set; the enhanced interference set I~\tilde I is used locally by its holder during calibration, and the server receives only the resulting model updates.

In our sample-level adaptation of FedEraser, the interference-holding client continues to participate in calibration and performs these operations through its control over local data. This requires neither additional server-side request fields nor adversarial capabilities beyond the stated threat model.


AuthorsOpenReview ↗

Response to Reviewer 5x8V (3/5): Q2 & Q3 - Target Screening and Sample Availability

Q2: Target Screening and Evaluation Eligibility

Targets are screened before attack construction using a separately trained and frozen surrogate model, and the subsequent eligibility check does not trigger target replacement or attack-set reconstruction. Therefore, this procedure does not introduce selection bias arising from iterative target filtering driven by the pre-unlearning check.

Target screening before attack construction. Specifically, we first screen candidate targets using the surrogate model and retain those it classifies correctly. We then construct the support and interference sets for the selected targets and perform federated training.

Eligibility check after federated training. After federated training, we check whether the resulting pre-unlearning global model correctly classifies each target. This check establishes eligibility for evaluating a successful flip from a correct prediction before unlearning to an incorrect prediction afterward. It is distinct from the surrogate-based screening performed before attack construction. The eligibility-check outcome does not trigger a return to attack construction, target replacement, or reconstruction of the attack sets.

In the revised manuscript, we revise the opening paragraph of Section C.4 to state this sequence explicitly and distinguish surrogate-based target screening before attack construction from the eligibility check after federated training.


Q3: Ensuring Sufficient Samples in the Selected Competing Class

We ensure sufficient samples in the selected competing class by enforcing a stronger input constraint before retrieval. Parameter configurations that violate this constraint are rejected before attack construction.

Sufficient input constraint. Specifically, we require

$ p_i \leq \left\lceil \frac{k_i}{C-1} \right\rceil, \qquad k_i \leq |\mathcal C_{\mathrm{int}}|, $

where CC is the total number of classes and Cint\mathcal C_{\mathrm{int}} is the candidate pool after excluding the target class. Here, pip_i denotes the number of distinct interference images selected, before any repetition used to form training records.

The kik_i retrieved neighbors belong to at most C−1C-1 competing classes. By the pigeonhole principle, the most frequent class contains at least ⌈ki/(C−1)⌉\lceil k_i/(C-1)\rceil samples. The strengthened constraint therefore guarantees that the selected class contains at least pip_i samples.

Experimental configurations. All reported experiments satisfy this condition. Both datasets have C=10C=10 classes:

DatasetRetrieved neighbors kik_iSelected images pip_iGuaranteed minimum count in the selected class
MNIST5056
CIFAR-101,00060112

In the revised manuscript, we explicitly state this sufficient condition in Algorithm 1 and specify that the candidate pool excludes the target class before retrieval. We validate these constraints before retrieval to ensure that the subsequent class-specific selection is feasible.


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

This paper investigates a security vulnerability in Federated Unlearning, where malicious clients can exploit unlearning requests to manipulate model predictions while preserving overall model utility. To address this problem, the authors propose BURST, an attack that uses public auxiliary data to construct two sets, a target-support set that reinforces correct predictions and an interference set that introduces competing-class influence. During unlearning, BURST removes the support set and strengthens the interference set to induce misclassification of previously correctly classified targets. The authors conduct experiments on MNIST and CIFAR-10, comparing BURST with existing attacks and internal baselines to evaluate attack effectiveness, global model utility, and computational efficiency. The reported results show that BURST achieves higher attack success rates than the compared methods while maintaining relatively limited global accuracy degradation and low attack-construction overhead.

Critical strengths
  • The paper highlights an important distinction between aggregate model utility and integrity on selected examples.
  • The method is easy to follow with promising effectiveness and measured construction efficiency.
Critical weaknesses

Major Weaknesses

  • The actual unlearning operation and the causal contribution of the proposed mechanism are unclear to me. The attacker controls both deletion requests and reconstruction-time data or updates, but the manuscript does not provide a clear FedEraser instantiation explaining how support records are removed while interference remains available for enhancement.
  • Another critical weakness is that the experimental reporting is insufficiently auditable to support the conclusions. The manuscript does not report the candidate counts, excluded trials, or uncertainty estimates required to interpret the headline ASRs. It also does not report key model, optimization, perturbation, and sample-budget values.

Minor Weaknesses

  • The paper does not examine how attack effectiveness changes with the support-set size, interference-set size, perturbation bounds, or optimization steps.
  • Equation (7) selects the competing class based on its frequency among the nearest auxiliary samples. However, the most frequent class in the surrogate feature space is not necessarily the class that can most strongly change the target prediction.
  • BURST selects auxiliary samples using cosine similarity in Equation (4), but Equations (5) and (9) optimize squared Euclidean distance between feature representations. These measures do not necessarily rank samples consistently when feature vectors have different magnitudes. The paper does not explain whether the representations are normalized or evaluate how this choice affects the attack.
Questions for authors
  • What exact FedEraser implementation or adaptation is used? Does the request remove individual support records or an entire client?
  • Please provide the actual model architectures, surrogate-training configuration, representation layer, optimization settings, sample counts, perturbation bounds and units, client assignments, and FU calibration parameters.
  • Please provide the results for the additional FU methods and data partitions claimed in the Introduction and Conclusion.
6 comments
AuthorsOpenReview ↗

Response to Reviewer ZxhR (1/4): Major W1 & Q1 - FedEraser Adaptation

Major Weakness 1 & Q1: Sample-Level FedEraser Adaptation and Interference Incorporation

In our sample-level FedEraser instantiation, the deletion request removes individual support records SS, not an entire client, while the original interference samples II, held by another malicious client, remain outside the scope of deletion. During local calibration, this client replaces II with the optimized I~\tilde I, whose influence enters reconstruction through the resulting model updates.

Incorporation into FedEraser. We implement BURST in a sample-level FedEraser setting, where deletion requests target specified data records and the affected clients continue to participate in local calibration using their remaining data. The procedure is as follows:

  1. The malicious client holding the support set SS requests its deletion and participates in local calibration using the remaining data.

  2. The interference set II, held by another malicious client, is outside the scope of this deletion request. During local calibration, this client replaces its existing II with the optimized I~\tilde I, preserves the labels, and computes model updates using the established local training procedure.

  3. The optimized interference samples thus influence unlearning reconstruction through the model updates produced by local calibration. The server receives only model updates and performs the established calibration and aggregation operations.

Consistency with the threat model. This procedure uses two capabilities specified in the threat model: issuing deletion requests for specified local records and controlling the malicious clients’ own local calibration data. Deleting SS corresponds to the former capability, while replacing II with I~\tilde I corresponds to the latter.

Interference enhancement takes place locally on the client, without submitting additional datasets to the server or adding fields to the deletion request. BURST therefore exploits malicious clients’ control over their local inputs during calibration in this sample-level setting. These operations require neither additional adversarial capabilities nor changes to the server-side calibration and aggregation procedures.

In the revised manuscript, we explicitly describe the sample-level adaptation of FedEraser and distinguish support-set deletion from client-side interference enhancement in the text and Algorithm 1.

AuthorsOpenReview ↗

Response to Reviewer ZxhR (2/4): Major W2 & Q2 - Settings and ASR Reporting

Major Weakness 2 & Q2: Experimental Configurations, Sample Budgets, and ASR Reporting

We report candidate-target counts, evaluation coverage, and uncertainty estimates alongside the headline ASRs, and provide the model, optimization, perturbation, and sample-budget settings in the appendix. We distinguish candidate targets from auxiliary-image pools, and distinct samples from repeated training records.

Candidate-target counts and uncertainty. All 22 MNIST candidate targets and all 35 CIFAR-10 candidate targets are evaluated, with none skipped. The table reports the success counts, ASR denominators, and 95% Wilson confidence intervals.

DatasetCandidate targetsEvaluated targetsSkipped targetsSuccessful attacksASR (%)95% Wilson CI (%)
MNIST222201463.64[42.95, 80.27]
CIFAR-10353503291.43[77.62, 97.04]

Models and training. In the default configuration, both datasets use 10 clients, a Dirichlet partition with α=1\alpha=1, and FedAvg. Local training uses SGD with cross-entropy loss, one epoch per round, a batch size of 128, momentum of 0.9, and weight decay of 0.0005. Surrogate parameters are frozen during attack construction.

ConfigurationMNISTCIFAR-10
ModelCNN2CIFAR-adapted ResNet-18
Representation dimension128512
Representation layerAfter the hidden-layer ReLU, before the final dropout and classifierAfter global average pooling, before the classifier
Surrogate training rounds2050
Surrogate learning rate0.0050.01
Federated training rounds2050
Federated training learning rate0.0050.01

Attack construction. The support and interference sets are constructed from the auxiliary-image pools below. Perturbations are optimized using Adam for 100 steps on MNIST and 200 steps on CIFAR-10, with learning rates of 0.01 and 0.005, respectively. Perturbation bounds are defined in unnormalized [0,1][0,1] pixel space.

ConfigurationMNISTCIFAR-10
Auxiliary datasetQMNISTSTL-10
Support candidate-pool size1,000800
Support nearest neighbors retrieved50800
Distinct support images selected560
Interference candidate-pool size1,0005,000
Interference nearest neighbors retrieved501,000
Distinct interference images selected560
Perturbation optimizerAdamAdam
Perturbation optimization steps100200
Perturbation optimization learning rate0.010.005
L∞L_\infty perturbation bound0.38/2558/255

Sample budgets and client assignments. In the default configuration and the corresponding baseline comparisons, all attack methods follow the same repetition rule: each distinct sample is included 100 times for MNIST and once for CIFAR-10. For BURST, the resulting budgets and client assignments are as follows:

ConfigurationMNISTCIFAR-10
Distinct support images560
Distinct interference images560
Total distinct auxiliary images10120
Training records per distinct image1001
Support training records50060
Interference training records50060
Total injected training records1,000120
Support removalClient 6 deletes 500 support recordsClient 6 deletes 60 support records
Interference enhancementClient 7 enhances 500 interference recordsClient 7 enhances 60 interference records

Federated unlearning. The sample-level FedEraser adaptation uses the following calibration settings:

ConfigurationMNISTCIFAR-10
Calibration interval Δt\Delta t52
Calibration learning rate0.0050.01
Local epochs per calibration11
AuthorsOpenReview ↗

Response to Reviewer ZxhR (3/4): Minor W1 & W2 - Sensitivity and Class Selection

Minor Weakness 1: Sensitivity to Set Sizes, Perturbation Bounds, and Perturbation Optimization Steps

We evaluate BURST’s sensitivity to support-set size, interference-set size, perturbation bounds, and perturbation optimization steps on MNIST. The results show that increasing set sizes or optimization steps does not necessarily improve ASR, while larger perturbation bounds yield higher ASR over the tested range.

Support-set and interference-set sizes. We examine support and interference sizes of 125, 250, 500, and 1,000 records in separate sweeps. Each set contains five distinct source images; the record count varies through repetition. These experiments therefore measure sensitivity to the number of training records associated with each set.

Records in the varied setSupport-size sweep: ASR (%)Interference-size sweep: ASR (%)
12568.1850.00
25059.0950.00
500 (default)63.6463.64
1,00063.6459.09

For support-set size, ASR ranges from 59.09% to 68.18%, with the highest observed value at 125 records. Increasing the support budget from 500 to 1,000 records leaves ASR unchanged at 63.64%. Thus, a larger support set does not necessarily produce a stronger attack.

For interference-set size, ASR increases from 50.00% at 125 and 250 records to 63.64% at 500 records, then decreases to 59.09% at 1,000 records. The default interference budget of 500 records achieves the highest ASR among the tested settings.

Perturbation bounds. We evaluate L∞L_\infty perturbation bounds of ϵ∈{0.05,0.10,0.20,0.30}\epsilon \in \{0.05, 0.10, 0.20, 0.30\}, measured in [0,1][0,1] pixel space before mean–standard deviation normalization.

Perturbation bound ϵ\epsilonASR (%)
0.0550.00
0.1050.00
0.2063.64
0.30 (default)68.18

ASR increases from 50.00% at ϵ=0.05\epsilon=0.05 and 0.100.10 to 63.64% at 0.200.20 and 68.18% at 0.300.30. These results characterize the relationship between the permitted perturbation magnitude and attack effectiveness: the attack succeeds on half of the evaluated targets at the two smallest tested bounds, while larger bounds increase the observed success rate.

Perturbation optimization steps. We examine perturbation optimization budgets of 10, 25, 50, 75, and 100 steps.

Perturbation optimization stepsASR (%)
1054.55
2554.55
5063.64
7559.09
100 (default)63.64

ASR increases from 54.55% at 10 and 25 steps to 63.64% at 50 steps. It reaches 59.09% at 75 steps and 63.64% at 100 steps.


Minor Weakness 2: Competing-Class Selection

Equation (7) uses class frequency as a low-cost heuristic to select a nearby, relatively well-represented, and label-consistent group of auxiliary samples in the surrogate feature space. It does not require the selected class to maximize the change in the target prediction.

Specifically, the nearest-neighbor retrieval identifies auxiliary samples close to the target in the surrogate feature space, while the frequency criterion selects a competing class with relatively abundant representation among these neighbors. This provides a coherent candidate pool for constructing the interference set without separately evaluating an attack for every possible competing class. The criterion therefore serves as a practical selection rule, with no claim of optimality.

With this selection rule, BURST achieves ASRs of 63.64% on MNIST and 91.43% on CIFAR-10. These results support the effectiveness of BURST using the proposed heuristic. In the revised manuscript, we clarify that Equation (7) selects a locally prevalent competing class to guide interference construction, rather than estimating which class produces the largest change in the target prediction.

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

This paper presents BURST, a malicious federated unlearning attack that combines target-support removal and interference enhancement to manipulate predictions on selected samples. The proposed approach constructs two complementary sets from public auxiliary data and coordinates their effects across federated training and unlearning. The paper reports a strong attack success rate of 91.43% on CIFAR-10, with a global accuracy reduction of 5.66 percentage points.

Critical strengths

The paper investigates a meaningful security problem involving client-controlled unlearning requests. The proposed mechanism is conceptually coherent and provides a targeted perspective on prediction integrity. The use of public auxiliary data and feature-based sample construction is practically interesting, and the experimental results demonstrate promising effectiveness and attack-construction efficiency. The inclusion of existing attack baselines and internal ablations is also appreciated.

Critical weaknesses

I have two major concerns.

First, the attack relies on removing the support set while introducing optimized interference samples during the unlearning reconstruction or calibration stage. However, the paper does not clearly explain how these additional interference inputs are incorporated into the standard FedEraser procedure, or whether the implementation requires modifications to the original protocol. This is a critical concern because the practical relevance of the attack depends on whether the adversary can perform these operations under the stated threat model. Without this clarification, it is difficult to determine whether BURST exposes a vulnerability inherent to the federated unlearning interface or relies on additional capabilities granted by the experimental implementation.

Second, BURST introduces both support and interference samples during federated training and subsequently removes the support set. However, the experiments do not include an exact-retraining reference after support removal, making it difficult to distinguish the effects of legitimate removal of previously injected data from additional manipulation during unlearning. Moreover, the S-only and I-only ablations differ from the full method in their training data composition, rather than isolating the unlearning operations while keeping the pre-unlearning model fixed. Consequently, although the reported attack success rates are promising, the current evidence does not fully establish the claimed complementary effects of support removal and interference enhancement. This limitation directly affects the interpretation of the paper's central technical contribution.

Questions for authors
  1. How exactly are the optimized interference samples incorporated into FedEraser's reconstruction or calibration procedure? Does the attack require modifications to the original protocol or additional adversarial capabilities beyond those specified in the threat model?
  2. What is the attack success rate when the model is retrained from scratch after removing the support set while retaining the original interference samples?
  3. Can the authors clarify or provide results that isolate support removal and interference enhancement using the same pre-unlearning model?
4 comments
AuthorsOpenReview ↗

Response to Reviewer kgiQ (1/2): W1 & Q1 - FedEraser and Threat Model

W1 & Q1: Incorporating Optimized Interference Samples into FedEraser within the Stated Threat Model

BURST introduces optimized interference samples through malicious clients’ existing local calibration steps in a sample-level FedEraser setting, without adding server interfaces or changing the server-side calibration and aggregation procedures. All required operations fall within the stated threat model.

Incorporation into FedEraser. We implement BURST in a sample-level FedEraser setting, where deletion requests target specified data records and the affected clients continue to participate in local calibration using their remaining data. The procedure is as follows:

  1. Support removal. Specifically, the malicious client holding the support set SS requests its deletion and participates in local calibration using the remaining data.

  2. Local interference enhancement. The interference set II, held by another malicious client, is outside the scope of this deletion request. During local calibration, this client replaces its existing II with the optimized I~\tilde I, preserves the labels, and computes model updates using the established local training procedure.

  3. Server-side reconstruction. The optimized interference samples thus influence unlearning reconstruction through the model updates produced by local calibration. The server receives only model updates and performs the established calibration and aggregation operations.

Consistency with the threat model. This procedure uses two capabilities specified in the threat model: issuing deletion requests for specified local records and controlling the malicious clients’ own local calibration data. Deleting SS corresponds to the former capability, while replacing II with I~\tilde I corresponds to the latter.

Interference enhancement takes place locally on the client, without submitting additional datasets to the server or adding fields to the deletion request. BURST therefore exploits malicious clients’ control over their local inputs during calibration in this sample-level setting.

In the revised manuscript and Algorithm 1, we explicitly describe the sample-level adaptation of FedEraser and distinguish support-set deletion from client-side interference enhancement.

AuthorsOpenReview ↗

Response to Reviewer kgiQ (2/2): W2, Q2 & Q3 - Retraining and Ablations

W2, Q2 & Q3: Exact Retraining and Fixed-Model Ablations of Support Removal and Interference Enhancement

We distinguish the effects of normal data deletion from those of interference enhancement during unlearning through an exact-retraining reference after support-set removal and operation-level ablations with a fixed pre-unlearning model. Across both datasets, full BURST achieves a higher attack success rate than either operation alone, supporting the complementary effects of support removal and interference enhancement.

Experimental setup. Specifically, we perform federated training on data containing both the support set SS and the original interference set II, and fix the resulting pre-unlearning model and training history. From this common starting point, we compare four reconstruction settings:

  • Replay: retains SS and the original II and executes the same reconstruction procedure.
  • Remove SS only: deletes SS while retaining the original II.
  • Enhance II only: retains SS and replaces II with the optimized I~\tilde I.
  • Full BURST: both deletes SS and replaces II with I~\tilde I.

These branches share the same training data composition and pre-unlearning model, differing only in the operations performed during reconstruction.

Additionally, Exact retraining removes SS from the training data, retains the original, unenhanced II, and reruns the complete federated training procedure from the initial model. This provides a reference for model behavior following normal support-set deletion.

Results. The attack success rates are as follows:

SettingSupport and interference operationsMNIST ASR (%)CIFAR-10 ASR (%)
ReplayRetain SS and the original II27.2741.18
Remove SS onlyDelete SS; retain the original II36.3664.71
Enhance II onlyRetain SS; replace II with I~\tilde I36.3652.94
Full BURSTDelete SS; replace II with I~\tilde I63.6494.12
Exact retrainingDelete SS; retain the original II; retrain from scratch4.5517.65

Exact-retraining reference (W2 & Q2). The exact-retraining reference shows that normal deletion of SS alone does not explain the attack effectiveness of full BURST. After removing SS and retaining the original II, exact retraining yields ASRs of 4.55% on MNIST and 17.65% on CIFAR-10, whereas full BURST achieves 63.64% and 94.12%, respectively.

Furthermore, under the same FedEraser reconstruction procedure, full BURST improves upon support removal alone by 27.28 and 29.41 percentage points, respectively, directly demonstrating the additional contribution of interference enhancement beyond support removal.

Complementarity under a fixed pre-unlearning model (W2 & Q3). The ablations with a fixed pre-unlearning model further support the complementarity of the two operations.

  • On MNIST, support removal alone and interference enhancement alone each achieve an ASR of 36.36%, while their combination reaches 63.64%.
  • On CIFAR-10, the individual operations achieve 64.71% and 52.94%, respectively, while their combination reaches 94.12%.

Thus, adding interference enhancement to support removal, or adding support removal to interference enhancement, further increases attack success. These comparisons attribute the combined benefit to the two operations during unlearning, rather than to differences in pre-unlearning models arising from different training data compositions.

Thank you for the detailed rebuttal and additional experiments. I appreciate the clarification of how optimized interference samples enter the sample-level FedEraser procedure through local calibration, and the exact-retraining reference and ablations using the same pre-unlearning model.

These additions substantially strengthen the experimental validation and largely address my main concerns. Some limitations remain regarding the scope of the sample-level adaptation and the reporting of uncertainty in the new ablation results, but these do not outweigh the improvements provided by the rebuttal.

Overall, I am updating my rating to Weak Accept.

Thank you for your careful follow-up and constructive feedback. We appreciate your recognition of the clarifications and additional experiments, as well as your time and effort throughout the review process.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary: Reviewer Feedback and Main Revisions

Dear Area Chair,

We sincerely thank you for the time and effort you devote to overseeing the review of our paper. For your convenience, we summarize the reviewer feedback, main revisions, and additional experimental evidence below.

Reviewer feedback

  • Reviewer kgiQ raises the rating to Weak Accept (Reviewer comment), highlighting the local-calibration clarification, fixed-model ablations, and exact-retraining reference.

  • Reviewer ZxhR maintains a positive rating (Reviewer comment), finding the adaptation and role of interference clearer after reading the revision and responses.

  • Reviewer 5x8V confirms that our rebuttal has addressed the original concerns and strengthened the experimental evidence (Reviewer comment). During the rebuttal period, the reviewer requests three additional clarifications on the finer details of reporting: (i) ASR comparability, (ii) threat-model assumptions and applicability, and (iii) runtime scope.

    Our November 14 (AoE) replies address all three in detail, as linked below. With the discussion period drawing to a close, we are awaiting further feedback from Reviewer 5x8V on these replies.

Main revisions and supporting evidence

1. How interference enters unlearning (Sections 3.2 and 4.3; Algorithm 1).

One cooperating client requests deletion of its support records; the other controls its local calibration data and returns the model updates. Algorithm 1 distinguishes the deletion request from data that remain local. The findings apply to unlearning procedures that permit both operations. Threat model and runtime clarification

2. Contributions of the two operations (Section 5.3; Figure 3).

We train with both support and original interference sets, then fix the pre-unlearning model and training history. Four reconstruction branches vary only support removal and interference enhancement; Replay applies neither. Exact retraining removes support, retains original interference, and reruns federated training from the initial model. All five settings use the same eligible targets. Values are successful flips / eligible targets.

SettingMNISTCIFAR-10
Replay6/2214/34
Support removal only8/2222/34
Interference enhancement only8/2218/34
Full BURST14/2232/34
Exact retraining1/226/34

Each operation outperforms Replay; their combination gives the highest success count on both datasets, supporting their complementary contributions. Exact retraining quantifies prediction changes from support deletion when training starts from scratch. Ablation response

3. Evaluation beyond the default setting (Section 5.5 and Appendix E).

  • KNOT: ASRs are 85.71% on MNIST and 81.82% on CIFAR-10, extending the evaluation to clustered unlearning. Test accuracy changes from 98.11% to 95.87% and from 81.61% to 75.98%, respectively.
  • Data heterogeneity: For Dirichlet alpha values of 1, 10, and 100, ASRs are 63.64%, 50.00%, and 52.38% on MNIST, and 91.43%, 84.00%, and 73.33% on CIFAR-10 (Figure 5). KNOT and heterogeneity response
  • Existing defenses: Training-only ASRs on MNIST are 75.00%, 63.64%, and 60.00% for Krum, Trimmed Mean, and FLAME. Unlearning-only ASRs are 72.73% and 68.18% for Trimmed Mean and FLAME. Figure 4 reports all five settings and final test accuracy. Defense response
  • Sensitivity and normalization: Appendix E covers support and interference record counts, perturbation bounds, optimization steps, and feature normalization. Larger record counts or more steps do not consistently increase ASR. Unnormalized and L2-normalized features both yield 15/22 successful flips (68.18%). Hyperparameter response

4. ASR comparability and runtime scope.

All experiments use the same candidate lists: 22 MNIST and 35 CIFAR-10 targets. Eligibility requires a correct pre-unlearning prediction. Our follow-up provides success counts, eligible-target denominators, and evaluation settings for Table 1 and the additional experiments.

Section 5.4 and Appendix C.6 limit the runtime comparison and efficiency conclusions to per-target attack construction with a prepared surrogate, excluding surrogate training, federated training, and unlearning execution.

Thank you again for your time and consideration. We hope this summary is helpful in your assessment of our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.