acceptodds
Under review as a conference paper at ICLR 2027

Active Unlearning Probing: Label Inference Attack from the Client's Vantage in Federated Unlearning

Abstract

Federated unlearning allows a client to request that its data be retracted from a jointly trained model, after which the server broadcasts the updated model to all participants. Prior forgotten-label inference typically assumes an attacker with a server vantage that can read individual client updates, yet in practice the server is usually trusted, so we ask whether the same inference attack can be mounted from the vantage of an ordinary client. Such a threat model is less restrictive and easier to realize. We show that the answer is yes. We propose Active Unlearning Probing (AUP), which issues short unlearning probes as a legitimate client to build a per-class gradient template, then matches the classifier-layer direction of the observed global-model change against these templates to infer which label class another user has forgotten. Using only its own labeled training data, with no shadow set, no labeled test set, and no queries to the victim, AUP identifies the forgotten label at 0.96 accuracy on CIFAR-10 and 0.80 on MNIST, likewise identifies labels on CIFAR-100 with a ViT-B/16 and on the Yahoo Answers text dataset, recovers three simultaneously deleted classes at an intersection-over-union (IoU) of 0.58, and remains effective under two different unlearning mechanisms. We also find that AUP becomes more effective as the federation grows, a phase transition governed by the signal-to-noise ratio rising with the number of clients. Code is available at https://anonymous.4open.science/r/AUP.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Unlearning, Label Inference Attack, Federated Learning, Machine Unlearning, Privacy Leakage

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

This paper studies privacy leakage in Federated Unlearning, asking whether an ordinary client can infer another client's forgotten labels without accessing private data or individual model updates. To address this, the authors propose Active Unlearning Probing (AUP), which builds class-specific unlearning templates from the attacker's own data and matches them to changes in the global model's classifier layer using cosine similarity. Experiments across datasets, model architectures, unlearning mechanisms, and federation sizes assess the attack's effectiveness, robustness, and scalability. Results show that AUP achieves high forgotten-label inference accuracy over the tested baselines, and its effectiveness generally improves as the number of clients increases.

Critical strengths
  • The paper identifies an important privacy risk without requiring access to the victim’s individual updates.
  • The experiments provide strong initial evidence across several datasets and show the value of focusing on the classifier layer.
Critical weaknesses

Major Weaknesses

  • My main concern is about the authors' claim that the victim’s signal remains fixed while background noise decreases as the number of clients grows. However, keeping the number of participating clients per round fixed does not guarantee that the victim’s update stays the same.
  • The reported accuracy measures label identification only for detected events. The paper does not provide equally detailed results for missed events, false alarms, or the full detection-and-identification process. It also omits simple comparison methods with the same information access, such as classifier-bias ranking, one-step gradient matching, and class-wise loss-change scoring on the attacker’s own data.

Minor Weaknesses

  • The main experiments assume IID data and the same label set across clients (Section 3). However, real federated learning systems often contain significant differences in client data distributions.
  • Section 3 assumes that the victim participates continuously during a multi-round unlearning campaign, while other clients are sampled normally. However, the paper does not evaluate situations where the victim participates discontinuously, where its unlearning stream is shorter, or where multiple clients request unlearning at overlapping times.
Questions for authors
  • Which algorithm was implemented? Did the reported experiments use the additional template-mean subtraction shown in Figure 1, or the version without this step in Algorithm 1 and Equation (8)?
  • For each setting, how many forget events were attempted, detected, and correctly identified? Please report detection recall, false-positive rates on blocks without unlearning, and the fraction of all events that were both detected and correctly identified.
  • What justifies neglecting prediction–feature covariance in Proposition 1? What assumptions control alignment over multiple ascent steps and rounds?
  • How does AUP compare with classifier-bias-only ranking, one-step class-gradient matching, and normalized class-wise loss-change scoring using the attacker’s own data?
7 comments
AuthorsOpenReview ↗

Reply to Reviewer zuPt [Part 1/5]

We thank the reviewer for the detailed review and for the concrete requests for baselines and detection statistics. Revisions are highlighted in red in the revised PDF. Additional results are in Appendix B, with the alignment analysis in Appendix C.


W1. The claim that the victim's signal stays fixed while background noise decreases as the number of clients grows

A1. Holding clients per round fixed does not hold the victim's update fixed, and the claim that the attack strengthens with federation size has been withdrawn.

We measured the quantities that enter the signal-to-noise ratio. On CIFAR-10, five seeds, twelve forget events per seed:

PartitionCCSamples/clientVictim local update normBackground noise sdIdentification
Fixed total data (IID)1050000.00100.00030.417
3016660.27160.00600.933
5010000.36610.00491.000
1005000.22550.01020.883
Fixed 500-sample shard105000.10900.00580.983
305000.17490.00300.900
505000.19060.00340.967
1005000.21600.00600.900

Under fixed-total-data partitioning the victim's update norm ranges from 0.00100.0010 to 0.36610.3661, more than two orders of magnitude, so it is not a fixed quantity, and the background noise is not monotone in CC under either protocol.

The decisive comparison is the controlled one at C=10C=10, where the client count is held fixed and only the per-client data volume changes: the per-class loss on the forget set before unlearning rises from 3.0×10−53.0\times 10^{-5} to 7.5×10−47.5\times 10^{-4}, the victim's local update norm rises from 0.00100.0010 to 0.10900.1090, about a hundredfold, and identification moves from 0.4170.417 to 0.9830.983. With a fixed per-client shard, identification across C=10,30,50,100C=10,30,50,100 is 0.983,0.900,0.967,0.9000.983, 0.900, 0.967, 0.900, a range of 0.0830.083 and not monotone.

The mechanism is the fitting state of the forgotten class. Gradient-ascent unlearning moves the classifier row in proportion to (p−y)(p-y). Under the fixed-total-data partition, a smaller federation gives each client more data, the model fits the forget class to near-zero loss, the ascent gradient nearly vanishes, and the victim's forgetting leaves almost no trace in the parameters.

The paper now states that attack effectiveness follows the magnitude of the forgetting update relative to the background noise, both of which the partition protocol and the fitting state determine (Section 5.4). Proposition 2 is stated as a sufficient condition at a fixed alignment.


W2 & Q2. Detection recall, false alarms, and the full detection-and-identification process

A2. Every stage is reported for the six monitoring settings, with block counts. Definitions are in Section 3: Rdet=TP/(TP+FN)R_{\mathrm{det}}=\mathrm{TP}/(\mathrm{TP}+\mathrm{FN}), FPR=FP/(FP+TN)\mathrm{FPR}=\mathrm{FP}/(\mathrm{FP}+\mathrm{TN}) on blocks without unlearning, AcondA_{\mathrm{cond}} on detected events, and ASRjoint=Rdet×Acond\mathrm{ASR_{joint}} = R_{\mathrm{det}} \times A_{\mathrm{cond}} over all attempted events. The CIFAR-10 settings use 60 events and 240 event-free blocks pooled over seeds, the multi-class setting 60 and 90, and the remaining settings 36 and 54. At k=3k=3, the attempted, detected, and correctly identified event counts are:

SettingAttemptedMagnitude: detected / correctProduct: detected / correct
CIFAR-10 / ResNet-186052 / 5259 / 59
CIFAR-10 / SGA-EWC6052 / 5160 / 59
CIFAR-10, C=100C=100600 / 07 / 7
CIFAR-100 / ViT-B/163628 / 2836 / 36
Yahoo Answers / MiniLM3636 / 3636 / 36
CIFAR-10, L=3L=36060 / 3360 / 33
AuthorsOpenReview ↗

Reply to Reviewer zuPt [Part 2/5]

We report both a threshold-free measure of separability and the operating points that different detection constants realize. Magnitude is the detector of Algorithm 3 applied to ∥Π(δ)∥\lVert \Pi (\delta )\rVert. Product applies the same threshold rule to ∥Π(δ)∥⋅max⁡ccos⁡(o,Tc)\lVert \Pi (\delta )\rVert\cdot \max_c \cos(o,T_c). Each operating-point cell is RdetR_{\mathrm{det}} / FPR / ASRjoint\mathrm{ASR}_{\mathrm{joint}}.

SettingScoreAUCk=1k=1k=2k=2k=3k=3
CIFAR-10 / ResNet-18magnitude0.9991.000 / 0.050 / 1.0000.983 / 0.017 / 0.9830.867 / 0.000 / 0.867
product0.9971.000 / 0.150 / 1.0001.000 / 0.087 / 1.0000.983 / 0.042 / 0.983
CIFAR-10 / SGA-EWCmagnitude0.9981.000 / 0.042 / 0.9830.983 / 0.013 / 0.9670.867 / 0.004 / 0.850
product0.9981.000 / 0.150 / 0.9831.000 / 0.096 / 0.9831.000 / 0.050 / 0.983
CIFAR-10, C=100C=100magnitude0.6960.100 / 0.017 / 0.0670.033 / 0.004 / 0.0330.000 / 0.000 / 0.000
product0.9410.483 / 0.029 / 0.4500.250 / 0.017 / 0.2330.117 / 0.008 / 0.117
CIFAR-100 / ViT-B/16magnitude0.9951.000 / 0.204 / 1.0001.000 / 0.093 / 1.0000.778 / 0.000 / 0.778
product0.9981.000 / 0.519 / 1.0001.000 / 0.352 / 1.0001.000 / 0.241 / 1.000
Yahoo Answers / MiniLMmagnitude1.0001.000 / 0.963 / 1.0001.000 / 0.907 / 1.0001.000 / 0.870 / 1.000
product1.0001.000 / 0.907 / 1.0001.000 / 0.907 / 1.0001.000 / 0.889 / 1.000
CIFAR-10, L=3L=3magnitude0.9421.000 / 0.467 / 0.5501.000 / 0.400 / 0.5501.000 / 0.389 / 0.550
product0.9611.000 / 0.456 / 0.5501.000 / 0.411 / 0.5501.000 / 0.389 / 0.550

The template-product AUC is between 0.9410.941 and 1.0001.000 in every monitoring setting, and detection performance depends on both score separability and threshold placement. Yahoo Answers has an AUC of 1.0001.000 while k=3k=3 realizes an FPR of 0.8700.870. The cause is where a threshold fixed on the B0B_0 calibration blocks falls relative to each setting's ambient scale. The mean ∥Π(δ)∥\lVert \Pi (\delta )\rVert over event-free monitoring blocks, first half and second half, against the calibration mean that sets the threshold, is 0.01390.0139 and 0.01450.0145 against 0.01300.0130 on CIFAR-10 (τ=0.0333\tau =0.0333), 0.17640.1764 and 0.11160.1116 against 0.16780.1678 on CIFAR-100 (τ=0.2344\tau =0.2344), and 0.10510.1051 and 0.13870.1387 against 0.03950.0395 on Yahoo Answers (τ=0.0759\tau =0.0759). On Yahoo Answers the ambient scale is 2.7 to 3.5 times the calibration mean, so the threshold sits below almost every event-free block. The detection constant is a setting-specific choice, and Section 5.2 reports the FPR next to every joint success rate.

At C=100C=100 the magnitude score has AUC 0.6960.696 and its k=3k=3 threshold fires on no block, while the template product has AUC 0.9410.941 and reaches ASRjoint=0.450\mathrm{ASR}_{\mathrm{joint}}=0.450 at FPR 0.0290.029 for k=1k=1. Incorporating class direction therefore improves event detection in this configuration.

AuthorsOpenReview ↗

Reply to Reviewer zuPt [Part 3/5]

W2 & Q4. Comparison with classifier-bias ranking, one-step class-gradient matching, and class-wise loss-change scoring

A3. We ran all three under the attacker's exact information access. Bias ranking orders classes by −Δbc-\Delta b_c and row-norm ranking by ∥ΔWc∥\lVert \Delta W_c\rVert. Neither needs any probing. One-step gradient matching compares oo with the vectorized classifier-layer gradient of the loss on the attacker's own class-cc samples. Loss and accuracy scoring rank classes by the change in the attacker's own per-class loss and accuracy. For the normalized form, we report three variants of the loss-change decoder below: the raw rise ℓc−ℓˉc\ell_c - \bar\ell_c, the relative rise (ℓc−ℓˉc)/ℓˉc(\ell_c - \bar\ell_c)/\bar\ell_c, and the calibration-standardized rise (ℓc−ℓˉc)/sc(\ell_c - \bar\ell_c)/s_c, where ℓˉc\bar\ell_c and scs_c are the mean and standard deviation of the attacker's class-cc loss over the B0B_0 calibration blocks. The table reports the raw form, and the normalized forms follow.

SettingBiasRow normOne-step gradLossAccuracyAUP
CIFAR-10 / ResNet-181.0000.8831.0001.0001.0001.000
CIFAR-10 / SGA-EWC1.0000.9000.9831.0001.0000.983
CIFAR-10, C=100C=1000.8670.5330.9000.8830.9500.933
CIFAR-100 / ViT-B/161.0001.0001.0000.8060.6671.000
Yahoo Answers / MiniLM1.0001.0001.0000.8610.7221.000
CIFAR-10, long-tailed (imb. 100)1.0000.9170.9170.4440.3891.000
CIFAR-10, L=3L=30.8000.0670.6000.7830.7670.550
Dirichlet α=0.5\alpha =0.50.2780.2500.3060.2500.1940.333

Without a defense, bias ranking matches AUP at 1.0001.000 on CIFAR-10, CIFAR-100, Yahoo Answers, and the long-tailed partition, and it requires no probing. On multi-class forgetting, bias, loss-change, and accuracy-change scoring reach 0.8000.800, 0.7830.783, and 0.7670.767, compared with 0.5500.550 for AUP. Under SGA-EWC these three decoders reach 1.0001.000 compared with 0.9830.983 for AUP. At C=100C=100, accuracy-change scoring reaches 0.9500.950 compared with 0.9330.933 for AUP. Under Dirichlet α=0.5\alpha =0.5, AUP reaches 0.3330.333 and one-step gradient matching 0.3060.306. Section 5.2 states the comparison in these terms.

Normalizing the loss-change decoder helps it, and we report where it then overtakes AUP:

SettingB3 rawB3 relativeB3 standardizedAUP
CIFAR-10 / ResNet-181.0001.0001.0001.000
CIFAR-10 / SGA-EWC1.0001.0001.0000.983
CIFAR-10, C=100C=1000.8830.9830.9500.933
CIFAR-100 / ViT-B/160.8060.8890.9721.000
Yahoo Answers / MiniLM0.8610.9440.7501.000
CIFAR-10, long-tailed (imb. 100)0.4440.5830.5561.000
CIFAR-10, L=3L=30.7830.6500.5170.550
Dirichlet α=0.5\alpha =0.50.2500.2780.3330.333

Relative normalization raises the loss-change decoder from 0.8830.883 to 0.9830.983 at C=100C=100, above AUP's 0.9330.933.

The comparison changes once the broadcast is perturbed, the defense that Section 6 proposes and Section 5.2 quantifies. Adding Gaussian noise of standard deviation σ\sigma to the broadcast classifier layer on CIFAR-100 / ViT-B/16, with pre-unlearning test accuracy in the first column:

σ\sigmaModel utilityBiasRow normLoss changeAccuracy changeOne-step gradAUP
00.8951.0001.0000.8060.6671.0001.000
0.0010.8970.0000.0280.8060.6391.0001.000
0.010.7970.0000.0830.3890.2781.0001.000
0.020.6910.0000.0560.0830.2500.5830.583

At σ=0.001\sigma =0.001, test accuracy is 0.8970.897 versus 0.8950.895 without noise. Bias ranking falls to 0.0000.000 and row-norm ranking to 0.0280.028, while AUP and one-step gradient matching retain 1.0001.000 identification accuracy. At σ=0.01\sigma =0.01 the two template-matching methods retain 1.0001.000, where test accuracy decreases by 9.89.8 percentage points and loss-change and accuracy-change scoring have fallen to 0.3890.389 and 0.2780.278. Both template matchers reach 0.5830.583 at σ=0.02\sigma =0.02, where the decrease is 20.420.4 percentage points.

Normalization does not change this. Under the same noise sweep on CIFAR-100 / ViT-B/16, the loss-change decoder reaches 0.806/0.889/0.9720.806 / 0.889 / 0.972 (raw / relative / standardized) at σ=0.001\sigma =0.001, and 0.389/0.250/0.3060.389 / 0.250 / 0.306 at σ=0.01\sigma =0.01, where AUP and one-step gradient matching are both still at 1.0001.000.

Reviewer_6ATYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether an ordinary federated learning client can infer which class another client has requested to forget, without observing individual client updates. The proposed Active Unlearning Probing (AUP) attack constructs class-specific unlearning templates using the attacker's own data and matches them against classifier-layer changes in the aggregated global model. The evaluation covers image and text classification, two gradient-based unlearning mechanisms, and both single-class and multi-class forgetting.

Critical strengths
  • Unlike prior forgotten-label inference attacks that rely on individual client updates, AUP operates using aggregated global model and the client attacker's own training data. This also exposes a privacy risk that remains even when individual updates are hidden.
  • The attack is simple and effective, particularly the deployable local variant, which constructs class-specific templates without requiring server rollback or actual forgetting requests.
  • The experiment is comprehensive, covering image and text classification, different unlearning mechanisms, and settings with varying numbers of classes.
Critical weaknesses
  • The theoretical justification for template alignment is not fully convincing. Proposition 1 neglects the covariance between predicted probabilities and features, which is not necessarily negligible under IID sampling. Moreover, the analysis focuses on a single gradient ascent step, while AUP involves multiple steps. Would the update directions remain aligned as the model changes throughout the forgetting process? The authors should clarify under what conditions this assumption holds.
  • The reported attack accuracy does not fully reflect the success of the overall attack. Section 4.3 explicitly evaluates forgotten-label identification conditional on successful event detection, although event detection is itself an essential part of AUP. Does the 0.96 accuracy reported in Table 2 account for forgetting events missed by the detector? I think the authors should also clarify how often normal training updates are incorrectly flagged, as both factors directly affect the practical effectiveness of the attack.
  • AUP assumes that client data follow an IID distribution, but non-IID data are common in real-world federated learning. I think it is important to consider how the method could be adapted to such settings. Moreover, even under IID sampling, some classes may have far fewer samples than others. Would the attack become less effective for these tail classes, and how can we ensure reliable inference in such cases?
  • Some notations are not clearly defined. For example, Eq. (1) introduces F(w)F(w), Fi(w)F_i(w), DiD_i, and DD without explicitly explaining their meanings. Although these are standard notations in federated learning, I think a brief clarification would improve readability.
Questions for authors
  • Would the update directions remain aligned as the model changes throughout the forgetting process?
  • I think the authors should also clarify how often normal training updates are incorrectly flagged, as both factors directly affect the practical effectiveness of the attack.
  • Would the attack become less effective for these tail classes, and how can we ensure reliable inference in such cases?
8 comments
AuthorsOpenReview ↗

Reply to Reviewer 6ATY [Part 1/2]

We thank the reviewer for the detailed reading of the theory and for the questions on heterogeneity. Revisions are highlighted in red in the revised PDF. Additional results are in Appendix B, with the alignment analysis in Appendix C.


W1 & Q1. Proposition 1 neglects the covariance between predicted probabilities and features, and the analysis covers a single ascent step while AUP uses several

A1. Proposition 1 and its proof have been rewritten to keep the covariance term and to state the alignment as a large-sample result with explicit conditions.

The covariance term is now kept. For the cross-entropy gradient on classifier row WkW_k, conditioning on y=cy = c and writing μc=E[ϕ(x)∣y=c]\mu_c = \mathbb{E}[\phi (x)\mid y = c] and pˉk,c=E[pk∣y=c]\bar p_{k,c} = \mathbb{E}[p_k\mid y = c], the exact population gradient is

gWk,c=(pˉk,c−1[k=c]) μc⊤+Cov⁡(pk, ϕ(x)∣y=c)⊤g_{W_k,c} = \big(\bar p_{k,c} - \mathbf{1}[k=c]\big)\,\mu_c^{\top} + \operatorname{Cov}\big(p_k,\,\phi(x)\mid y=c\big)^{\top}

so the classifier-layer direction is not a scalar multiple of the mean feature. The key point is that the attacker and the victim estimate the same population quantity gcg_c, so the covariance term is common to both and does not need to be small for the directions to agree. What remains is estimation error. Writing uc=gc/∥gc∥u_c = g_c/\lVert g_c\rVert for the population direction and P⊥=I−ucuc⊤P_\perp = I - u_c u_c^{\top} for the projection orthogonal to it, a delta-method expansion gives

E[cos⁡(g^cA,g^cB)]=1−tr⁡(P⊥Σc)2∥gc∥2(1nA+1nB)+o(1nA+1nB)\mathbb{E}\big[\cos(\hat g_c^{A},\hat g_c^{B})\big] = 1 - \frac{\operatorname{tr}(P_\perp\Sigma_c)}{2\lVert g_c\rVert^{2}}\Big(\frac{1}{n_A}+\frac{1}{n_B}\Big) + o\Big(\frac{1}{n_A}+\frac{1}{n_B}\Big)

Only the gradient dispersion orthogonal to the population direction enters: noise along ucu_c rescales both estimates without rotating either, so it leaves the cosine unchanged. The constant is tr⁡(P⊥Σc)/∥gc∥2\operatorname{tr}(P_\perp \Sigma_c)/\lVert g_c\rVert^{2}, and the alignment tends to one as both sample sizes grow.

We now label this a large-sample result and state the conditions it requires:

ConditionStatus
Per-sample gradients have finite second momentsAssumed
∥gc∥\lVert g_c\rVert bounded away from zeroAssumed. The measured minimum-to-median per-class template norm ratio is 0.784 on CIFAR-10, 0.909 on Yahoo Answers, and 0.156 on CIFAR-100 with a ViT, and no template exists for a class the attacker does not hold
Both parties differentiate at the same parameter wwApproximate in deployment: the attacker probes at the broadcast state it holds, while the victim's campaign starts from a later state. The alignment that remains is measured below
Independent, disjoint data splitsHolds, since the attacker and victim hold disjoint shards. Identical distribution holds under the IID partition

Across multiple steps the direction is stable within a local campaign and decays across federation rounds. We measure both granularities on CIFAR-10, three seeds and all ten classes:

GranularityStep/round 1FinalCosine with the first direction
Local SGD steps inside one unlearning campaign (10 steps)0.8810.9130.926
Federation rounds inside one event block (5 rounds)0.9150.4840.515

Within a local campaign the cumulative update does not rotate away: its cosine with the class template rises from 0.881 to 0.913, and its cosine with the first step stays at 0.926 after ten steps. The decay comes from the federated rounds, where the broadcast model moves and other clients' updates accumulate. Proposition 1 concerns gradients at a common model state. The measurements above report how the class-gradient direction changes along the unlearning trajectory.

A direct measurement of the quantity in the proposition shows where it saturates. The cosine between the victim's local unlearning update and the attacker's template, measured as the attacker's per-class probe size ncn_c grows:

ncn_c5102050100
cos⁡\cos(victim update, attacker template)0.7240.7750.7960.7540.786

The cosine is 0.724 at nc=5n_c = 5 and between 0.754 and 0.796 for nc=10n_c = 10 to 100, so increasing the probe budget beyond ten samples changes it little, and the remaining gap to one comes from the different model states above (Appendix C).

AuthorsOpenReview ↗

Reply to Reviewer 6ATY [Part 2/2]

W2 & Q2. The reported accuracy is conditional on detection. Does 0.96 account for missed events, and how often are normal updates flagged

A2. The 0.96 in Table 2 is identification on detected events and does not account for missed events. We now separate and report both stages. Section 3 defines ASRjoint=Rdet×Acond\mathrm{ASR_{joint}} = R_{\mathrm{det}} \times A_{\mathrm{cond}} over all attempted events, so missed events are charged against the attack.

On the main CIFAR-10 configuration, with 60 unlearning events and 240 event-free blocks pooled over seeds and Acond=1.00A_{\mathrm{cond}} = 1.00:

ScoreRdetR_{\mathrm{det}}Event-free blocks flaggedFPRASRjoint\mathrm{ASR_{joint}}
Magnitude, k=3k = 3 (Algorithm 3)0.8670 of 2400.0000.867
Magnitude, k=1k = 11.00012 of 2400.0501.000
Template product, k=3k = 30.98310 of 2400.0420.983

So on CIFAR-10 normal training updates are flagged in none of 240 event-free blocks at k=3k = 3, and end-to-end success over all attempted events is 0.867.

Across settings the template-product AUC is between 0.941 and 1.000, and detection performance depends on both score separability and threshold placement. Yahoo Answers has an AUC of 1.000 while k=3k = 3 realizes an FPR of 0.870, and at C=100C = 100 the magnitude score has AUC 0.696 while the template product has 0.941. Appendix B reports the AUC, the k=3k = 3 block counts, and the k=1,2,3k = 1, 2, 3 sweep for every monitoring setting, and Section 5.2 reports the FPR next to every joint success rate.


W3 & Q3. Non-IID data, and tail classes with few samples

A3. We separate two effects that a Dirichlet partition mixes together: heterogeneity among the honest peers, and heterogeneity of the attacker's own shard.

Setting (C=50C = 50, CIFAR-10)Attacker's shardIdentification
Dirichlet α=0.5\alpha = 0.5, attacker and victim on IID shardscomplete, ≈100 per class0.583
Dirichlet α=0.1\alpha = 0.1, attacker and victim on IID shardscomplete, ≈100 per class0.389
Dirichlet α=0.5\alpha = 0.5, attacker also skewedall 10 classes, min 6 samples0.333
Dirichlet α=0.1\alpha = 0.1, attacker also skewed4 classes absent0.278

We group the α=0.1\alpha = 0.1 events by the attacker's sample count for the forgotten class:

Attacker's samples of the forgotten classEventsCorrectIdentification
01400.000
1–201350.385
>20950.556

Every event whose forgotten class is absent from the attacker's shard fails, since no template can be constructed, and accuracy over the remaining events rises monotonically with the available sample count. Class coverage is therefore a precondition the attacker can check on its own shard before probing. When the class is present, identification on the main configuration is 1.000 at per-class probe budgets of nc=5n_c = 5, 2020, and 100100 (Table 9).

On a long-tailed CIFAR-10 partition, where class frequencies follow an imbalance factor instead of a Dirichlet draw, AUP reaches 0.944 identification accuracy at an imbalance factor of 10 and 1.000 at 100.

Template transfer therefore requires per-class coverage on the attacker's own shard, and it degrades as the honest peers become more heterogeneous. Section 3 now states coverage as an explicit precondition alongside the IID premise, and the Future Work section lists template estimation under incomplete label coverage as the next step.


W4. Notation in Eq. (1) is not defined

A4. Section 3 now defines DiD_i as the local dataset of client ii, D=⋃i=1CDiD = \bigcup_{i=1}^{C} D_i as the union over clients, Fi(w)F_i(w) as the local empirical risk, and F(w)F(w) as the global data-weighted empirical risk, immediately after Eq. (1).


We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

Thanks, and these changes might be included in the revised version of the manuscript.

Reviewer_PX6cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AUP, an attack that allows a federated client to infer another client's forgotten labels from global model updates. The method builds class-specific templates using the attacker's own data. Experiments show high inference accuracy across several datasets and models.

Critical strengths

The paper studies label inference from an ordinary client's perspective, without access to other clients' updates. Unlike previous server-side attacks, AUP only uses the broadcast global model and the attacker's own data. This extends the threat model of existing work and reveals a potential privacy risk even when individual client updates are protected.

Critical weaknesses

The evaluation does not fully establish the effectiveness of AUP. Section 4.3 states that label inference accuracy is reported only when an unlearning event has been detected. However, the paper does not evaluate the detection rate or false-positive rate. Since event detection is a necessary step of the attack, the reported accuracy alone does not show how reliably AUP works in practice. A complete evaluation of the detection and inference process is needed to support the paper's main claims.

Questions for authors

Section 4.3 reports label inference accuracy conditional on detecting an unlearning event. Could the authors clarify how well the event detector performs, particularly its detection rate and false-positive rate? How would missed events affect the overall attack success rate?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer PX6c [Part 1/2]

We thank the reviewer for the focused question on event detection. Revisions are highlighted in red in the revised PDF. Additional results are in Appendix B, with the alignment analysis in Appendix C.


W1 & Q1. Detection rate, false-positive rate, and the effect of missed events on overall attack success

A1. We now evaluate detection and identification separately and jointly, with the quantities defined in Section 3:

  • detection recall Rdet=TP/(TP+FN)R_{\mathrm{det}}=\mathrm{TP}/(\mathrm{TP}+\mathrm{FN}),
  • false-positive rate FPR=FP/(FP+TN)\mathrm{FPR}=\mathrm{FP}/(\mathrm{FP}+\mathrm{TN}) on blocks containing no unlearning,
  • conditional identification accuracy AcondA_{\mathrm{cond}} on detected events,
  • end-to-end success ASRjoint=Rdet×Acond\mathrm{ASR_{joint}} = R_{\mathrm{det}} \times A_{\mathrm{cond}} over all attempted events.

ASRjoint\mathrm{ASR_{joint}} counts every attempted event, including missed ones. On the main CIFAR-10 configuration, monitoring streams contain 60 unlearning events and 240 event-free blocks pooled over seeds, and identification on detected events reaches 1.001.00, so the joint rate equals the detection recall.

The table below reports both a threshold-free measure of separability and the operating points that different values of the detection constant kk realize. Magnitude is the detector of Algorithm 3, applying τ=mean+kσbase\tau =\mathrm{mean}+k\sigma_{\mathrm{base}} to ∥Π(δ)∥\lVert \Pi (\delta )\rVert. Product applies the same threshold rule to ∥Π(δ)∥⋅max⁡ccos⁡(o,Tc)\lVert \Pi (\delta )\rVert\cdot \max_c \cos(o,T_c). Each operating-point cell is RdetR_{\mathrm{det}} / FPR / ASRjoint\mathrm{ASR_{joint}}.

SettingScoreAUCk=1k=1k=2k=2k=3k=3
CIFAR-10 / ResNet-18magnitude0.9991.000 / 0.050 / 1.0000.983 / 0.017 / 0.9830.867 / 0.000 / 0.867
product0.9971.000 / 0.150 / 1.0001.000 / 0.087 / 1.0000.983 / 0.042 / 0.983
CIFAR-10 / SGA-EWCmagnitude0.9981.000 / 0.042 / 0.9830.983 / 0.013 / 0.9670.867 / 0.004 / 0.850
product0.9981.000 / 0.150 / 0.9831.000 / 0.096 / 0.9831.000 / 0.050 / 0.983
CIFAR-10, C=100C=100magnitude0.6960.100 / 0.017 / 0.0670.033 / 0.004 / 0.0330.000 / 0.000 / 0.000
product0.9410.483 / 0.029 / 0.4500.250 / 0.017 / 0.2330.117 / 0.008 / 0.117
CIFAR-100 / ViT-B/16magnitude0.9951.000 / 0.204 / 1.0001.000 / 0.093 / 1.0000.778 / 0.000 / 0.778
product0.9981.000 / 0.519 / 1.0001.000 / 0.352 / 1.0001.000 / 0.241 / 1.000
Yahoo Answers / MiniLMmagnitude1.0001.000 / 0.963 / 1.0001.000 / 0.907 / 1.0001.000 / 0.870 / 1.000
product1.0001.000 / 0.907 / 1.0001.000 / 0.907 / 1.0001.000 / 0.889 / 1.000
CIFAR-10, L=3L=3magnitude0.9421.000 / 0.467 / 0.5501.000 / 0.400 / 0.5501.000 / 0.389 / 0.550
product0.9611.000 / 0.456 / 0.5501.000 / 0.411 / 0.5501.000 / 0.389 / 0.550

Three findings follow.

(i) The unlearning event is separable in every setting. The template-product AUC ranges from 0.9410.941 to 1.0001.000, so the signal an ordinary client observes is present throughout. On CIFAR-10 the complete attack reaches ASRjoint=0.867\mathrm{ASR_{joint}}=0.867 with zero false alarms among 240 event-free blocks at k=3k=3, and 1.0001.000 at k=1k=1 with an FPR of 0.0500.050.

(ii) Detection performance depends on both score separability and threshold placement. On Yahoo Answers the AUC is 1.0001.000, yet k=3k=3 flags 47 of 54 event-free blocks, an FPR of 0.8700.870. The threshold is fixed from the B0B_0 calibration blocks that precede the monitoring stream, and the ambient scale of an event-free block does not stay at the calibration level. Mean ∥Π(δ)∥\lVert \Pi (\delta )\rVert over each group of blocks, pooled over seeds:

SettingCalibration blocksEvent-free, first halfEvent-free, second halfτ\tau at k=3k=3FPR at k=3k=3
CIFAR-10 / ResNet-180.01300.01390.01450.03330.000
CIFAR-10 / SGA-EWC0.01310.01380.01450.03350.004
CIFAR-100 / ViT-B/160.16780.17640.11160.23440.000
Yahoo Answers / MiniLM0.03950.10510.13870.07590.870
AuthorsOpenReview ↗

Reply to Reviewer PX6c [Part 2/2]

The threshold column is the seed-mean of mean+3σbase\mathrm{mean}+3\sigma_{\mathrm{base}}, and thresholds are applied per seed. On the three vision settings the event-free magnitude stays at or below the calibration scale, so τ\tau is reached in at most one block out of 240. On Yahoo Answers it is 2.7 times the calibration mean over the first half of the monitoring stream and 3.5 times over the second, so τ\tau sits below almost every event-free block. The detection constant is a calibration choice tied to each setting's background distribution, and the false-positive rates we report are empirical.

(iii) At C=100C=100 the template-product score retains separability that the magnitude score loses. The magnitude AUC is 0.6960.696 and its k=3k=3 threshold fires on no block, while the template product has AUC 0.9410.941 and reaches ASRjoint=0.450\mathrm{ASR}_{\mathrm{joint}}=0.450 at FPR 0.0290.029 for k=1k=1. Incorporating class direction therefore improves event detection in this configuration.


The detector reads only preceding blocks: Δbaseline\Delta_{\mathrm{baseline}} and σbase\sigma_{\mathrm{base}} are estimated on B0B_0 calibration blocks that precede the monitoring stream and are disjoint from the blocks on which FPR is measured, and the event labels used to score RdetR_{\mathrm{det}} and FPR are evaluation ground truth unavailable to the attacker at monitoring time. Appendix A states the protocol.

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up to Reviewer PX6c

Dear Reviewer PX6c,

Thank you again for reviewing our paper. As the discussion period draws to a close, we summarize how the revision addresses your comment on the event detector.

The revised PDF evaluates the detector directly: detection recall, false-positive rate, and end-to-end success are defined in Section 3 and reported for every monitoring setting in Section 5.2 and Appendix B (Tables 6 and 7). On the main CIFAR-10 configuration, the detector at k=3k=3 raises no false alarm over 240 event-free blocks, and end-to-end success over all attempted events is 0.867.

We would be grateful if you could take a look at our response and the revised version. If these results address your concern, we hope they can be reflected in your assessment.

Best regards, Authors of Submission 20830

AuthorsOpenReview ↗

A further note before the discussion closes

Dear Reviewer PX6c,

Apologies for taking another moment of your time, and thank you once again for your careful and constructive review.

With about ten hours left in the discussion period, we would be sincerely grateful if you could let us know whether our response addressed your concern about the event detector, or whether any question still remains.

Best regards, Authors of Submission 20830

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion for the Area Chair

Dear Area Chair,

Thank you and the reviewers for your time. We summarize the evidence added during the discussion and its current status.

  1. End-to-end evaluation of the detection stage (raised by all three reviewers). The revision defines detection recall, false-positive rate, conditional identification, and end-to-end success (Section 3) and reports them for every monitoring setting with block counts (Section 5.2, Appendix B, Tables 6 and 7). On the main CIFAR-10 configuration the detector at k=3k=3 raises no false alarm over 240 event-free blocks and end-to-end success over all attempted events is 0.867.

  2. Federation-size claim withdrawn and replaced by a controlled measurement (Reviewer zuPt). The victim's update norm and the background noise are now measured directly (Appendix B, Table 11). The original claim that the attack strengthens with the number of clients is removed from the abstract, introduction, Section 5.4, and the conclusion. A controlled comparison at C=10C=10 attributes the effect to per-client data volume and the fitting state of the forgotten class.

  3. Same-information client-side baselines (Reviewer zuPt). Bias ranking, row-norm ranking, one-step gradient matching, and loss- and accuracy-change scoring, including three normalized loss-change variants, are reported (Appendix B, Tables 8 and 9). Without a defense, bias ranking ties AUP in several settings and exceeds it on multi-class forgetting. Under broadcast noise of σ=0.001\sigma=0.001, which leaves test accuracy unchanged, the probe-free rankings fall to chance while AUP and one-step gradient matching retain 1.000 (Section 5.2).

  4. Proposition 1 rederived with the covariance term (Reviewers 6ATY and zuPt). The alignment result now keeps the prediction–feature covariance and gives the leading-order expansion in terms of the dispersion orthogonal to the population gradient, with step-level and round-level alignment measured (Appendix C).

  5. Heterogeneity, class coverage, and participation (Reviewers 6ATY and zuPt). Two Dirichlet regimes are separated (Section 6), identification is broken down by the attacker's per-class sample count, and victim participation, campaign length, and concurrent requests are swept (Appendix B, Table 10). Figure 1 was redrawn to match Algorithm 1.

Reviewers 6ATY and zuPt have confirmed that their concerns were addressed and updated their ratings. Reviewer PX6c's single concern is the detection evaluation in item 1. We have not yet received a response to our replies and follow-ups.

Thank you again for handling our submission.

Best regards, Authors of Submission 20830

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.