acceptodds
Under review as a conference paper at ICLR 2027

Certified Unlearning for Distilled Teacher-Student Models with a Black-Box Teacher

Abstract

We study certified machine unlearning for student models distilled from a teacher accessible only through black-box prediction queries. After deployment, the student must unlearn requested samples, while the teacher is unavailable for retraining and accessible only through black-box prediction queries. We introduce a student-side certification framework that re-distills the student on the retained data using only black-box teacher access, and releases a randomized model by carefully calibrating the perturbation to certify indistinguishability against the ideal but counterfactual retrain-from-scratch reference, where the latter would require retraining the (inaccessible) teacher without the forget samples. Our certificate builds on a student sensitivity bound for distillation, which we obtain by characterizing how the teacher's predictions would change under retraining without the forget samples, and then propagating this bound through the distillation objective to bound the student's parameter shift. The final student sensitivity is used to perturb the released student parameters with calibrated Gaussian noise to ensure the target certification guarantees. In doing so, our work provides a first study of unlearning certification for black-box teacher-student distillation. While our work is primarily theoretical and relies on several regularity assumptions for tractability, we hope our results will provide interesting insights for unlearning in this increasingly common regime.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, certified unlearning, knowledge distillation, distilled student-teacher models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_a23zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper defines certified unlearning for a distilled student against the ideal reference of retraining both teacher and student, and shows that simply re-distilling with the original teacher still leaks deleted data through soft labels. It proposes CKDU, which re-distills the student on retained data and adds Gaussian noise calibrated by combining a student perturbation bound with an SGD stability bound on the teacher.

Critical strengths
  1. The paper convincingly shows that re-distilling with the original teacher is not valid unlearning, and its pipeline-level definition is likely to be useful to the community.
  2. The theoretical chain is modular and appears correct, and the convex εTV\varepsilon_{TV} values in Table 4 can be reproduced from Eq. 10.
Critical weaknesses
  1. The certificate is non-vacuous only when the teacher is nearly uninformative or effectively frozen (soft labels in [0.456, 0.544] in the main experiments; a vacuous bound at the stated App. H step size), and it requires detailed knowledge of the teacher's training that contradicts the black-box setting.
  2. Several results are inconsistent with the paper's own formulas: Table 1 shows a 19-point gap between Redistill-only and Retrain although Fig. 2 implies near-identical models, and the noise in Tables 4 and 7 appears smaller than Eq. 11 requires.
Questions for authors
  1. How can Table 1 show a 19-point forget-accuracy gap when Fig. 2 bounds the logit difference to about 10−710^{-7}? An explanation could restore my Soundness score, while a failure of the BJB_J bound would lower my rating to 1.
  2. Which strong-convexity constant and teacher step schedule were used to compute the noise in Tables 4 and 7?
  3. What step size was actually used in App. H, and how far did the teacher's parameters move?
  4. Can you show a setting with an informative teacher, valid constants and reasonable utility? This would most likely raise my rating.
4 comments

Thank you for the insightful comments.

[W1, Q4] To directly investigate whether meaningful certified unlearning is possible with informative teachers, we revised the experiments of Section 6 in the revised paper with two settings where the teacher is confident in its predictions, its regularity constants are derived in closed form, and the Gaussian noise is calibrated by our bounds.

In both settings reported below, we use a non-convex tanh teacher and a non-convex student (other settings are reported in Section 6 of our revised paper). On MNIST (digits 3 and 8), the inputs are reduced to 2 dimensions by PCA, and on CIFAR-10 (all ten classes), the features of a frozen ViT-B/16 are reduced to 16 dimensions by PCA. In both cases the principal components are computed on public images that are not used for training. Importantly, the students learn mostly from the teacher's predictions (λ=0.9\lambda=0.9 on MNIST and λ=1\lambda=1 on CIFAR-10), allowing us to directly assess whether the teacher provides useful supervision. The table below shows the test accuracy of the teacher, the mean confidence of its predictions on the test data, and the test accuracy of the student released by CKDU at a forget ratio of 10%.

SetupTeacher accuracyTeacher confidenceStudent accuracy
MNIST (2D PCA)84.682.682.6
CIFAR-10 (16D PCA, 10 class)93.193.681.3

We also report the test accuracy, the forget accuracy on the deleted samples, the retain accuracy on the remaining samples, and the membership attack accuracy (MIA) on the released student. The results show that CKDU achieves similar utility to Retrain while providing a strong forgetting performance aligning with our theoretical findings. The values below are from Table 1 of the revised paper (non-convex student, forget ratio of 10%), which also reports Retrain + noise, Redistill-only, Labels-only, and the convex student.

MNIST (2D PCA)

MethodTestForgetRetainMIA
Original84.8 ±\pm 0.1100.0 ±\pm 0.0100.0 ±\pm 0.068.4 ±\pm 0.5
Retrain84.7 ±\pm 0.285.0 ±\pm 1.1100.0 ±\pm 0.053.4 ±\pm 0.4
CKDU82.6 ±\pm 1.580.0 ±\pm 0.595.0 ±\pm 0.951.7 ±\pm 0.1

CIFAR-10 (16D PCA)

MethodTestForgetRetainMIA
Original88.4 ±\pm 1.298.0 ±\pm 1.196.4 ±\pm 0.062.8 ±\pm 0.1
Retrain84.4 ±\pm 0.386.0 ±\pm 0.798.9 ±\pm 0.052.1 ±\pm 0.4
CKDU81.3 ±\pm 0.582.8 ±\pm 1.395.3 ±\pm 1.350.2 ±\pm 0.5

We provide the details of each experiment in Appendix G.

[W1] We thank the reviewer for raising this question. In our framework, black-box access means that the student cannot access or modify teacher's parameters, gradients, or internal states. Our certification guarantees instead rely on metadata about the teacher's training procedure, such as the training dataset size, learning-rate schedule, number of training steps, and valid upper bounds on the regularity constants, without requiring access to the teacher's internal parameters.

This information plays an important role in our theoretical analysis. Since the ideal retraining reference involves a counterfactual teacher trained without the forget samples, the original teacher's predictions alone cannot determine how its outputs would change under deletion. The training metadata enables us to bound this otherwise inaccessible discrepancy through the SGD stability analysis in Theorem 5.3, which is then propagated to the student through Theorem 5.1.

For third-party teachers, the teacher vendor can provide the required training metadata alongside the black-box prediction interface, which enables formal certification without disclosing the teacher's parameters or internal states. We discuss this setting in greater detail in our response to Reviewer ouZU comment [W2, Q4, Q1], including emerging regulatory precedents for sharing technical documentation between model providers and downstream developers.

[W2, Q1] The 19-point gap comes from a typo in Table 1 (which is now Table 16 in the revised paper). The CIFAR-10 forget accuracy of Redistill-only is 79.3%, not 89.3%, so the gap to Retrain (70.0%) is 9.3 points. We have corrected this in the revised paper. The BJB_J bound itself holds, since on these models the logit difference between the two students stays below the upper bound.

We also want to note that since the discrepancy plotted in Fig. 2 measures parameter-distance rather than classification accuracy, and since accuracy counts how many samples fall on the correct side of the decision boundary, even a small parameter or logit perturbation can change predicted classes for samples with very small classification margins. This helps explain why a 9.3-point gap in forget accuracy can coexist with a small parameter discrepancy, while the observed parameter distance remains within the derived bound.

[W2, Q2] Thanks for catching this. We noticed that this was due to a typo, the distillation coefficient λ=0.5\lambda=0.5 should be λ=0.3\lambda=0.3 instead, this is now fixed in the revised version. Now the noise in Table 4 (now Table 12 in the revised version) follows from Eq. 11 and Corollary 5.2 as expected.

For Table 4 (now Table 12 in the revised version), the strong-convexity constant is the ridge weight μ\mu reported in Table 2 (now Table 10 in the revised version), which is 3×10−73\times10^{-7} for XOR and 10−710^{-7} for MNIST and CIFAR-10. The teacher is trained by single-sample SGD for T=32,768T=32{,}768 steps with the step size αt=1/t\alpha_t=1/t on XOR and αt=0.3/(t+1000)\alpha_t=0.3/(t+1000) on MNIST and CIFAR-10 (Appendix E.2, now Appendix J.2 in the revised version). For Table 7 (now Table 15 in the revised version), no strong-convexity constant is used since the student for this setting is non-convex, and the noise is calibrated with the non-convex bound of Theorem 5.1.

Additionally, we have also carefully checked our paper for other typos. We have found that the reported configurations and constants for the nonconvex experiments were reflecting an earlier version of the experiment. We have now corrected the reported configurations and constants in Appendix J, the empirical accuracy results and figures were unchanged.

[Q3] We thank the reviewer for catching this typo. The step size printed in Appendix H (now Appendix M in the revised paper) had a typo, and we corrected it in the revised paper. As stated in Appendix M, the step size was chosen by screening the bound, which gives αt≈2×10−13\alpha_t\approx2\times10^{-13}, that is, ∑tαt≈2×10−10\sum_t\alpha_t\approx2\times10^{-10} over T=1,000T=1{,}000 steps. All results in Tables 10 and 11 (now Tables 19 and 20 in the revised paper) were computed using the correct step size.

Regarding parameter movement, each SGD step moves the teacher by at most αtL^\alpha_t\hat L, so over the T=1,000T=1{,}000 steps the teacher moves at most L^∑tαt\hat L\sum_t\alpha_t from its pretrained initialization, which is at most 1.2×10−71.2\times10^{-7}. However, the regularity constants of a ResNet-56 are large (L^\hat L between 410.9 and 567.6 in Table 20). Therefore, a small parameter change does not directly mean a small change in the loss, since the change in the loss can be up to L^\hat L times larger than the parameter change. This is why Theorem 5.3 accounts for every step through these constants instead of using the parameter distance, and why this experiment is an ablation and not one of our certified results. For a confident teacher with closed-form constants, we refer to our answer to W1 and Q4.

Thank you for the detailed response and the revised Section 6. The noise calibration now checks out: I can reproduce Delta in Tables 4/12 and 7/15, and sigma/Delta ~ 2.575 matches the analytic Gaussian mechanism. The App. M clarification and the metadata-based reading of "black-box" are also reasonable; the abstract should state the latter explicitly.

Two concerns remain about the new main experiments:

  1. Per App. G.2, the teacher is trained mostly on private data, and the phase that touches D_S has step sizes summing to only 1.47 (MNIST) and 9.5e-4 (CIFAR-10). The teacher is informative but effectively frozen with respect to the forget set. This is the W1 mechanism and should be stated in Section 6.

  2. Given (1), Redistill-only and Retrain should be nearly identical, yet Table 1 shows MIA gaps of about 13 points (e.g., 66.1 vs 53.1 on CIFAR-10). Also, with lambda = 1 on CIFAR-10 the student never sees true labels, yet retain accuracy is 100.0, above the teacher's 93.1%. Reporting prediction flips and logit differences between Redistill-only and Retrain, and how accuracy is computed, would resolve this.

We thank the reviewer for the careful follow-up and for independently verifying our sensitivity calculations and noise calibration. We have now clarified in the abstract (highlighted in red) that black-box access includes prediction access to the teacher, and our certification requires teacher training metadata, but does not require access to the teacher model parameters or dataset. We have also conducted additional measurements to answer the two remaining questions.

On teacher sensitivity to the forget set: Our main experiments use a two-phase teacher-training procedure (Appendix G.2): the teacher is first trained on teacher-only data and then further trained on the student dataset, which includes the forget samples. The second phase uses relatively small learning rates, particularly on CIFAR-10. We will clarify this aspect of the experimental setup in Section 6.

Importantly, the original and retrained teachers are not identical. Figure 2 already demonstrates that the teacher discrepancy on retained samples is nonzero and increases with the forget ratio. To evaluate this more directly, we additionally measured the total variation between the original and retrained teacher predictions in the convex CIFAR-10 experiment at 10% deletion. The average total variation is 9.6×10−49.6\times10^{-4} on the forget samples and 7.8×10−57.8\times10^{-5} on the retained samples, approximately 12 times larger on the former. Thus, removing the forget samples has a measurable effect on the teacher predictions, particularly on the forgotten examples. Our certification analysis explicitly accounts for this discrepancy rather than assuming that the teacher remains unchanged.

On the differences between Redistill-only and Retrain: We appreciate the reviewer raising this question and have further evaluated the two student models in the convex CIFAR-10 experiment highlighted by the reviewer. We obtained the root-mean-squared ℓ2\ell_2 difference between their logits as 4.8×10−54.8\times10^{-5} on the forget samples and 2.12×10−52.12\times10^{-5} on the test samples. Their top-1 predictions differ on 4% of the forget samples and only 0.2% of the test samples.

These results confirm that the two student models have similar predictions, consistent with the reviewer’s expectation. They also show that their differences are concentrated more strongly on the forget samples than on the test samples. The 4% prediction-flip rate is also consistent with the forget accuracies reported in Table 1 (88% for Retrain and 92% for Redistill-only).

The MIA, however, evaluates a different quantity: membership distinguishability based on per-example cross-entropy losses, rather than top-1 classification predictions. The attack uses randomly selected samples for threshold calibration and evaluation. Its balanced accuracy depends on the separation of the forget and test-loss distributions relative to the selected threshold. Consequently, high predictive agreement does not necessarily imply similar loss-threshold MIA scores. Our additional logit and prediction-flip measurements establish that the student predictions are closely aligned, while the MIA results characterize their behavior under a separate loss-based evaluation. In particular, the MIA gap should not be interpreted as a comparably large difference in the students’ classification behavior.

On accuracy computation: All accuracies in Table 1 are top-1 accuracies computed against the true labels. The previously reported 93.1% teacher accuracy refers to the held-out CIFAR-10 test set, whereas the student’s 100% retain accuracy is evaluated on its retained training samples. These are different evaluation sets. On the same held-out test set, the Redistill-only student achieves 85.2%, below the teacher’s 93.1%. Although the student is trained with teacher-generated soft labels when λ=1\lambda=1, this does not constrain it to reproduce the teacher’s classifications exactly.

We appreciate the reviewer’s careful evaluation and constructive discussion, which helped us clarify the experimental setting and interpretation of our results. We will also incorporate these clarifications into the paper.

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

The paper discusses certified unlearning for knowledge distillation, with the teacher model treated as a black-box model. By injecting calibrated Gaussian noise into a student model re-distilled on the retained data using the original teacher, the proposed method achieves (ϵ,δ+γ)(\epsilon, \delta + \gamma)-certified unlearning, making the student statistically indistinguishable from a clean re-train with the teacher model retrained first on the retained data and then the student re-distilled with the updated teacher. The formulation in this paper assumes that the teacher and student both use the same training inputs, with full visibility of the retained data and forget set during unlearning for both teacher and student, which narrows the overall setting compared to general knowledge distillation. Overall, I think the paper provides a detailed theoretical analysis and a derivation of the statistical bound under the specific teacher-student shared training input setting, providing good insights into certified unlearning for knowledge distillation, but the scope of the current setting is too narrow to be practical.

Critical strengths
  1. The theoretical analysis and derivation look correct and rigorous given the problem setting and formulation.

  2. The empirical results provide support for the certified unlearning effectiveness of the proposed CKDU, showing that the unlearned student model is close to the clean retraining reference for simplified architectures like one-hidden-layer tanh networks.

Critical weaknesses
  1. The assumption that the teacher and student share the same training input makes the scope of the paper narrower than that of general knowledge distillation, which allows the teacher and student to have very different training inputs. This is particularly relevant to tasks like local adaptation, where the student will have local data, but the teacher may be trained on a broader or different dataset. Thus, the proposed method should be extended beyond the teacher and student sharing the same input settings.

  2. The second weakness is related to the first but is about teacher model data accessibility. The current formulation assumes access to the entire teacher training dataset throughout the unlearning process, which is reasonable when the teacher and student share the same training data. However, it may not be realistic when the teacher and student are trained on different datasets. In such cases, a more practical setting would be one where only the forget set is available, without access to the remaining teacher training data. I think the proposed method should consider this setting when it is extended beyond teacher and student sharing the same dataset.

  3. The current method still requires the complete re-distillation of the student model, which could be computationally expensive. I think it would be sensible to check whether the current formulation can be combined with other approximate unlearning or certified unlearning methods and derive an updated relevant certification bound, allowing a computationally more efficient approach.

  4. The current bound works well for simplified architectures like the one-hidden-layer tanh network included in the main text. While experiments are done on deeper networks, the current methods do not provide a formal certification for such networks, as the network architectures do not give a closed-form solution to regularity constants. Consequently, the proposed method does not provide an implementable solution for formally certified unlearning for deep networks.

Questions for authors

As shown in my previous comments, my major concern about the work is the scope of the claimed knowledge distillation and the implementation of formally certified unlearning, which requires the authors to provide further work instead of clarity. The only question is:

  1. Could the authors provide further justification for the assumption that the teacher and student share the same training inputs? It would also be helpful to distinguish the theoretical results based on this assumption from those that do not rely on it. This could make some of the theoretical results general to the broader knowledge distillation settings.
7 comments

We thank the reviewer for the constructive comments, please see our responses below. We organized our responses such that [W1], [W2],... refers to the comments on weaknesses (in the same order), whereas [Q1] refers to the questions for authors. Our changes are highlighted in our revised manuscript in blue.

[W1, W2, Q1] We thank the reviewer for this insightful comment. As suggested by the reviewer, we have revised our framework to accommodate different teacher student datasets. Specifically, our framework can be extended to the setting when the teacher and student samples are from the same domain, but include possibly different samples.

In the revised paper, we now separate the teacher dataset DT\mathcal{D}_T and the student dataset DS\mathcal{D}_S in Section 3, where the forget set is given by DU⊆DT∩DSD_U\subseteq D_T\cap D_S. Beyond its own dataset DS\mathcal{D}_S, the student has no knowledge of the remaining teacher-only samples DT\DSD_T\backslash\mathcal{D}_S of the teacher. We then updated our theoretical derivations to reflect this new setup, which required applying the following changes to the theoretical results in Section 5,

  • Theorem 5.1 and Corollary 5.2 compare two student training runs on the same retained student data DR=DS∖DUD_R=D_S\setminus D_U, so they do not depend on the teacher's dataset.
  • Theorem 5.3 depends on the teacher's training and the number of deleted samples, but not on the student's data. Since it bounds the change of the teacher's predictions at every input, it also holds on student inputs that the teacher never saw during its training.

Therefore, the only implicit requirements are that the teacher can be queried on the student's inputs and that the deleted samples are in both datasets.

To demonstrate these updates, we show below our CIFAR-10 experiment with the non-convex student from Table 1 of our revised paper (which includes both convex and non-convex experiments). In this experiment, the teacher is trained on 21,000 CIFAR-10 images, while the student only has 1,000 of these images, and the 100 deleted images (forget ratio of 10%) are part of the student's data. The student's owner does not have access to the remaining teacher data. This also matches a setting suggested by the reviewer, where the student will have local data, but the teacher may be trained on a broader data.

MethodTestForgetRetainMIA
Original88.4 ±\pm 1.298.0 ±\pm 1.196.4 ±\pm 0.062.8 ±\pm 0.1
Retrain84.4 ±\pm 0.386.0 ±\pm 0.798.9 ±\pm 0.052.1 ±\pm 0.4
Retrain + noise83.1 ±\pm 1.684.0 ±\pm 1.197.7 ±\pm 0.251.1 ±\pm 0.5
CKDU81.3 ±\pm 0.582.8 ±\pm 1.395.3 ±\pm 1.350.2 ±\pm 0.5
Redistill-only84.7 ±\pm 1.394.0 ±\pm 1.396.7 ±\pm 0.159.1 ±\pm 0.2
Labels-only73.3 ±\pm 0.959.0 ±\pm 1.198.7 ±\pm 0.866.2 ±\pm 0.1

As shown in the table, CKDU stays close to Retrain + noise on every metric, and it reduces the MIA of the original student from 62.8% to 50.2% suggesting successful unlearning. On the other hand, the model achieved by Redistill-only still keeps 94.0% forget accuracy and 59.1% MIA, since the teacher's predictions still carry information about the deleted samples supporting the main claim of the paper. The details of this setup are presented in Appendix G in our revised version.

[W4] We thank the reviewer for this observation. Obtaining tight, computable certificates for unrestricted, end-to-end trained deep networks remains challenging. However, this limitation should be distinguished from the theoretical scope and contributions of our work. Our work is primarily a theoretical study of certified unlearning in a previously unexplored setting, rather than a claim to have solved practical certification for arbitrary deep architectures. Theoretical certified-unlearning guarantees necessarily operate under explicit mathematical assumptions or algorithmic restrictions. For example, foundational works by Guo et al. (2020), Sekhari et al. (2021), and Neel et al. (2021) establish their guarantees under convexity, strong convexity, smoothness, or related regularity conditions. The significance of such results lies in whether their conditions are well motivated and enable meaningful new guarantees, rather than whether they immediately yield practical certificates for unrestricted deep networks.

In our setting, these assumptions address a fundamental challenge specific to black-box distillation. Unlike conventional unlearning, here the correct retraining reference requires retraining an inaccessible teacher before redistilling the student. Without quantitative control over how the teacher's predictions change under deletion, this counterfactual reference cannot be meaningfully bounded from the available predictions alone. Our key theoretical contribution is to establish this missing connection, and a formal characterization of how otherwise inaccessible teacher influence propagates through distillation and can be controlled to certify student unlearning. To that end, Theorem 5.1 characterizes how deviations in teacher predictions propagate through finite-step student training, including smooth nonconvex distillation objectives under the stated regularity conditions, hence are not restricted to one-hidden-layer tanh networks. Corollary 5.2 derives tighter contraction-based bounds under strong convexity. Theorem 5.3 bounds the counterfactual teacher-output discrepancy through SGD stability, and Theorems 5.4-5.5 integrate these results into an end-to-end certified-unlearning guarantee against the counterfactual retraining reference. The architectures in Section 6 are chosen to rigorously instantiate and numerically validate these theoretical guarantees. Importantly, our theory requires valid bounds on regularity constants, not necessarily closed-form expressions, and separates teacher stability from student sensitivity, so we believe future bounds of either kind for broader deep networks can be plugged into CKDU without changing the underlying certification principle.

We therefore view practical certification for broader deep-network architectures as an important direction for future work, while distinguishing this open challenge from the theoretical characterization and guarantees established under our stated assumptions.

[W3] Following the reviewer's suggestions, we extended CKDU to checkpoint-based certified unlearning and derived the updated certification bound (Theorem 5.6, proof in Appendix F). Recall that in our original version, the student starts from a random initialization and took TST_S distillation steps on the retained data. Instead of complete re-distillation, we now allow the student to start from an already trained checkpoint TC<TST_C < T_S (trained during original training using the full student dataset, including the forget samples) and take only the remaining K=TS−TCK=T_S-T_C steps on the retained data where TST_S is the total training horizon (number of distillation steps) of the student.

The derivation in the revised version has two main components for the noise required to achieve certification in this warm-start case. The first one is the teacher discrepancy representing the impact of the original teacher, which is handled as in the analysis of CKDU. The second one covers the influence of the forget samples still stored in the checkpoint. Therefore, a later checkpoint takes fewer unlearning steps, but it needs more noise to erase the information about the forget samples stored during the initial training.

We evaluated this trade-off on MNIST 3 vs. 8 with two-dimensional inputs and the convex student for the checkpoints TC=16T_C=16, 24, and 32.

Methodσ\sigmaTestForgetRetainMIA
Original-84.798.0100.063.9
Retrain-83.785.0100.051.9
TC=16T_C=16
Redistill-only-84.693.0100.064.7
Retrain + noise1.7482.082.098.050.9
CKDU1.7481.982.598.150.3
TC=24T_C=24
Redistill-only-84.793.0100.065.4
Retrain + noise2.5280.779.696.051.5
CKDU2.5280.680.496.250.7
TC=32T_C=32
Redistill-only-84.798.0100.070.9
Retrain + noise5.5855.651.064.552.2
CKDU5.5856.154.062.751.7

The results demonstrate the trade-off of Theorem 5.6. Increasing the checkpoint clearly increases the noise level which is supported by the fact that the forget accuracy of Redistill-only increases from 93% to 98%. That is, the confidence on the forget samples is higher in the later checkpoints, which requires more noise to remove that information intuitively. With the bound we derived in the updated document, this influence is also carefully considered for certified unlearning.

We provide all details about this experiment in Appendix H.

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

This work studies the problem of certified unlearning. It provides formal guarantes between an unlearned model ( knowledge of some forget samples is removed from a trained model) and a model re-trained from scratch on the dataset without forget samples. Many deployed models are students distilled from larger teachers, thus even if we retrain students with new dataset, original teacher could still distill knowledge of deleted samples onto the student during distillation.

This work proposes Certified Knowledge Distillation Unlearning (CKDU) which consists of two steps. Let D_r denote the dataset without forget samples. First stage consists of re-distillation from the original teacher with this dataset DrD_r, resulting in w^r\hat{w}_r, which would still consist of knowledge regarding deleted samples. Next stage is to release a model with weights w^r+N(0,σ2I)\hat{w}_r + N(0, \sigma^2 I) with calibrated σ\sigma which results in a high-probability bound on distance between w^r\hat{w}_r and wr∗{w}^{*}_r (the student model trained with D_r with teacher that is traIned on D_r, thus leaking no knowledge about deleted samples).

While the proposed mechanism is simple, theoretical analysis is necessary to provide a certified guarantee on the unlearning procedure. This work provides the following analysis to provide ceritification guarantee on the released student model w^r+N(0,σ2I)\hat{w}_r + N(0, \sigma^2 I)

Theorem 5.1 (student side) bounds the deviation in student parameters when trained on two different teacher predictive distributions. It uses a gradient perturbation bounds on the KL term in the distillation objective. It results in deviation in T_s steps as deviation Δ\Delta related to a product of the teacher smoothness and total deviation between teacher distributions and a multiplicative term GTSG_{T_{S}}, which is expansive in the non-convex case and contractive in strongly convex case.

Theorem 5.3 (teacher side) bounds the total deviation, using the high-probablity adaptation of Hard et. al’s SGD stability arguments. It takes minimum of two bounds: (a) a deterministic tail bound after the event (“no forget sample yet”), and (b) a Chernoff bound on a weighted Bernoulli sum.

Finally these are combined with Gaussian mechanism to provide guarantees on the certified unlearning given that the noise σ\sigma used in the perturbation during release follows the constraints listed in the above results.

Empirical experiments on different setups compare the guarantees (which are computable for many setups in these tasks) to show that the CKDU framework provides reliable unlearning procedure.

Critical strengths
  • Easy to understand following the main arguments in the paper. The explanation behind redistill-on original teacher and the illustrative figure shows why perturbation is necessary for redistilled student on the original teacher.
  • Theorem 5.1 works on non-convex student loss functions, and provides a tighter bounds when the student loss is strongly convex
  • Easy to grasp proof structure for the theoretical analysis
  • Constants are derived in closed form for main experiments
Critical weaknesses
  • Weak threat model — how realistic is a setup where a regulatory body asks to delete samples from a student but not the teacher API.
  • Certificate of unlearning needs a lot of knowledge about how the teacher was trained (Assumption 4.1, Lipschitz and smoothness constants L and beta over all labels, step size, T, and training details like single-sample, w/ replacement SGD with no momentum, weight decay, augmentation, mini-batching, etc. ). This would be violated in any practical training recipe.
  • CKDU fully retrains the student per request, which could lead to inefficiency in the proposal. How are sequential delete requests are handled?
  • Missing baseline where no distillation from teacher is used? Why not just perform retraining on the new dataset without KD distillation term involving the teacher? At the scale, its very likely that KD is giving minimal returns.
  • Fig. 2 shows teacher bounds about 1e4× above observed values and student bounds about 1e5–1e6× above. The bounds are also nearly flat across forget ratios.
Questions for authors
  • Why is it that teacher can only be queries through black-box? Why this specific constraint? Can you access teacher intermediate outputs? You were able to access teacher’s intermediate outputs or any other hidden state during the initial KD phase? It is because this assumption simplifies the underlying theory?
  • How does CKDU handle subsequent delete requests? Would this method become prohibitively expensive at such requests?
  • What are the realistic setup where it makes sense for a student to unlearn a forget sample, but you would not expect the teacher to remove that forget sample’s knowledge? Since any privacy or regulatory concern would apply the same to both teacher and student w.r.t. the forget sample.
  • How would a practitioner get the constants? How would anyone obtain L, β, and the exact SGD protocol for a third-party teacher? Can the certificate tolerate minibatching, without-replacement shuffling, momentum, or Adam?
4 comments

We thank the reviewer for the thoughtful questions, please see our responses below. We organized our responses such that [W1], [W2],... refers to the comments on weaknesses (in the same order), whereas [Q1], [Q2],... refers to the questions for authors. Our changes are highlighted in our revised manuscript in blue.

[W1, Q3] Thank you for raising this important question. We agree that if a deletion request applies to all uses of a sample, then both the teacher and student may need to unlearn it. However, our work addresses a different, practically relevant setting, where the student's use of a sample must be discontinued, whereas the teacher may retain an independent authorization to use it. The key distinction is that data-use restrictions can be specific to a particular purpose or entity, rather than necessarily requiring universal deletion.

For a concrete deployment scenario, consider a general-purpose teacher model operated by an upstream provider and a smaller, task-specific student distilled and deployed by a downstream organization. The two models may serve different purposes and be subject to different data-use permissions. For instance, consider a healthcare provider authorized to retain patient information for clinical care, while a separate organization uses a distilled student model for research or planning. A patient may opt out of the latter use without preventing the healthcare provider from continuing to use the information for individual care. In such a scenario, the downstream organization must remove the influence of those samples from its student model, while the teacher's authorized use can still continue. As such, student-only unlearning is a meaningful operational requirement, rather than an arbitrary assumption that the teacher is exempt from privacy obligations.

Importantly, this distinction is supported by established data-protection principles and precedents:

  • The NHS National Data Opt-Out lets patients restrict the use of their data for research and planning without affecting its use for individual care (NHS England, 2019)
  • GDPR Art. 21(2)-(3) allows individuals to object to the use of their data for direct marketing, without prohibiting its processing for other lawful purposes (European Parliament and Council of the European Union, 2016)
  • In a Google Spain judgment (CJEU C-131/12; Court of Justice of the European Union, 2014), the Court established that a search engine could be required to remove links to personal information while the original webpage lawfully retains that information.

These precedents establish the asymmetric data deletion principle that motivates our work, in particular, the obligation to discontinue one use or dissemination of information does not automatically imply that every other authorized use must also cease. In our setting, the restricted downstream use corresponds to the student, while the authorized upstream use corresponds to the teacher.

Beyond the data-protection perspective, the asymmetric unlearning setup can also be useful for model maintenance, such as when the training dataset contains samples that are confusing for the student, for example samples not related to the student's task, or ambiguous samples close to the decision boundary. A downstream provider may wish to remove the influence of such samples from the student model, while retaining the general purpose teacher model, which can be used to train other student models for other tasks. While these use-cases for model repair are outside of the scope of our current work, they are interesting future directions.

In our revised paper, we added Appendix A to provide a detailed discussion of these motivations.

[W2, Q4, Q1] We do not access the teacher's intermediate outputs or hidden states, but rather metadata information about the teacher training procedure such as teacher's dataset size nTn_T, number of training steps TT, learning-rate schedule {αt}\lbrace \alpha_t\rbrace and valid upper bounds on the regularity constants LL and β\beta. These assumptions enable our formal certification guarantees, therefore their availability is important for practical deployment.

For a third-party teacher, a possible deployment scenario is that the teacher provider supplies the relevant metadata about the training process together with valid upper bounds on LL and β\beta, alongside the black-box prediction interface. Sharing model-related information across organizational boundaries is an emerging regulatory practice. For instance, the recent EU AI Act (Article 53) (European Parliament and Council of the European Union, 2024) establishes requirements for general-purpose AI model providers, such as sharing technical documentation for the training/testing process with downstream providers who intend to integrate these models into their AI systems, as downstream vendors cannot inspect the hidden weights or training pipelines of black-box models.

Importantly, our regularity constants can be derived analytically from the properties of the loss function and model architecture under suitable boundedness conditions, rather than requiring access to the particular trained parameter values. We demonstrate this in our controlled experiments (Sec. 6 and App. G), where the required constants are analytically derived. Obtaining rigorous bounds for arbitrary large-scale teacher architectures can be considerably more challenging. The role of these assumptions is to establish a bound on how the teacher's predictions could change under retraining without the forget samples. Since this counterfactual teacher is unavailable, some form of stability information is needed to provide a nontrivial theoretical certificate. Without any constraints on the teacher training procedure or its sensitivity to data deletion, the original teacher's black-box predictions alone generally cannot determine how a retrained teacher would behave. Such dependencies on regularity or stability assumptions are also common in existing certified unlearning literature, as discussed in our response to Reviewer hKTp, W4.

References

[Q4] No, these enhancements are currently out of our scope. We believe that our main procedure can be extended to other training recipes if, for instance, an appropriate teacher stability bound is established. In particular, our framework separates teacher-side stability from student-side sensitivity and certification. Specifically, Theorem 5.1 bounds student sensitivity in terms of the discrepancy between the original and retrained teacher predictions, independently of how the teacher was trained. Theorem 5.3 then provides one way to bound this discrepancy using SGD stability. Consequently, if an appropriate teacher-output discrepancy bound is established for another training procedure, it can be incorporated directly into the student sensitivity bound and subsequent Gaussian certification in Theorems 5.4-5.5. Therefore, extending CKDU to other training procedures may not require redesigning the entire certification mechanism, but rather establishing the corresponding teacher-side stability and student-side sensitivity guarantees for the training procedures of interest. While these are out of the scope of our current work, deriving these guarantees for minibatch SGD, without-replacement sampling, momentum, and adaptive optimizers would be an interesting direction for future work.

[W4] Following the reviewer's comment, we evaluated the student retraining baseline without distillation from the teacher (λ=0\lambda=0, labels-only retraining). We have considered this baseline (which we called "Labels-only") with both our old and new experiments.

Old experiments. We also evaluated Labels-only retraining in the convex-student setting of the original submission (DS=DT\mathcal D_S=\mathcal D_T, forget ratio of 20%, mean over three seeds).

DatasetMethodTestForgetRetainMIA
XORRetrain88.584.696.654.6
CKDU88.182.695.552.4
Labels-only88.585.396.253.4
MNISTRetrain96.095.4100.049.6
CKDU95.694.8100.049.9
Labels-only96.295.3100.051.3
CIFAR-10Retrain73.670.0100.050.3
CKDU73.570.5100.049.9
Labels-only75.773.3100.048.6

Since in the original experiments the student is trained on the same dataset as the teacher and the weight of the knowledge distillation term in the student's loss function was less than 0.5 for all three datasets, the effect of the teacher's guidance was relatively small with respect to the true labels. Therefore, in those experiments Labels-only retraining achieved similar utility with CKDU. However, the student is still trained to reproduce the distribution of the knowledge-distilled reference.

New experiments. In our revised paper, we have revised our framework and added new experiments to accommodate new reviewer comments. Specifically, we have extended our framework to accommodate mismatched teacher-student datasets in response to Reviewer 2 comment [W1, W2, Q1], and more confident teacher models in response to Reviewer 3 comment [W1, Q4]. In these experiments, the teacher and student are set up as described in Section 6 and Appendix G of our revised paper. We added our new results, which include the Labels-only training baseline at the student, to Table 1 of our revised paper (including both convex and non-convex students). The table below shows our CIFAR-10 results with the non-convex student.

MethodTestForgetRetainMIA
Original88.4 ±\pm 1.298.0 ±\pm 1.196.4 ±\pm 0.062.8 ±\pm 0.1
Retrain84.4 ±\pm 0.386.0 ±\pm 0.798.9 ±\pm 0.052.1 ±\pm 0.4
Retrain + noise83.1 ±\pm 1.684.0 ±\pm 1.197.7 ±\pm 0.251.1 ±\pm 0.5
CKDU81.3 ±\pm 0.582.8 ±\pm 1.395.3 ±\pm 1.350.2 ±\pm 0.5
Redistill-only84.7 ±\pm 1.394.0 ±\pm 1.396.7 ±\pm 0.159.1 ±\pm 0.2
Labels-only73.3 ±\pm 0.959.0 ±\pm 1.198.7 ±\pm 0.866.2 ±\pm 0.1

We observe that the utility of the labels-only retraining is much lower than the CKDU (73.3% against 81.3% test accuracy). By removing the teacher's guidance, we are giving up the utility advantage that a more capable and generic teacher can provide to a smaller and more task-specific student.

By comparing our new experiments with the old ones we can see the effect of the teacher's guidance on the student's utility significantly.

[W5] Thank you for this observation. We would like to clarify that our theoretical bounds are designed to account for worst-case sensitivity under the stated assumptions, rather than tightly estimate the discrepancies observed in a particular training run. Specifically, the teacher bound uses global regularity constants and high-probability SGD stability analysis, while the student bound propagates this discrepancy through the distillation trajectory. As a result, the bounds can be conservative compared with the observed discrepancies, and the discrepancy may vary depending on the experimental setup. For example, in our new experiments in Section 6 (Fig. 2 in the revised version), we observed substantially smaller gaps between the theoretical bounds and empirical observations, by about two orders of magnitude for the teacher and up to three orders for the student compared with the original experiments.

Regarding the apparent flatness across forget ratios in the original experiments, we would like to clarify that the logarithmic scale makes the growth less visually apparent. The bounds do increase with the forget ratio: from 5% to 20%, they grow by approximately 1.9×1.9\times on XOR and 3.7×3.7\times on MNIST and CIFAR-10, as reported in Table 12 of the revised paper.

While these bounds can be conservative, they provide the sensitivity guarantees needed for our Gaussian noise calibration. Developing tighter bounds that retain these guarantees is an interesting direction for future work.

[Q1] We study this black-box setting since real-world applications increasingly rely on multi-stage training pipelines where the teacher models are often proprietary models that another party owns and serves through a prediction API, often through budget-restricted limited query access. The party who distills the student by using teacher predictions but cannot access the teacher's parameters can therefore neither read nor modify it.

These constraints actually make the underlying theory and the approach to the problem harder and do not simplify it. With access to the teacher's parameters and training data, one could unlearn the teacher with an existing certified unlearning method and then redistill the student. The student would inherit the teacher's guarantee by post-processing, with no student-side analysis. However, in this work we instead study the worst-case discrepancy of the teacher predictions without having access to its parameters and analyze the sensitivity of the student parameters depending on this worst-case teacher stability analysis.

[W3, Q2] To improve the unlearning efficiency, in our revised version we extended our framework to accommodate a warm-start student model from an already trained checkpoint (rather than fully retraining the student), and taking fewer unlearning rounds. To that end, in the revised version we provide a warm-start sensitivity analysis of CKDU for strongly convex students (Theorem 5.6, proof in Appendix F, further discussion in this rebuttal is given under Reviewer hKTp, W3).

We then handle sequential requests by leveraging the warm-start analysis. Assume that the student has a checkpoint saved at a training step TCT_C (from the original training). For the sequential unlearning student can use the checkpoint as a starting point for every request, and the forget set can be updated to include all samples requested to be removed at that point. In that way the cost of the unlearning requests is reduced to KK steps representing remaining iterations after the checkpoint, instead of TST_S steps from scratch representing the whole student training iterations.

In our current setting, the released noisy models are never used as a starting point, so the noise of earlier releases does not accumulate when there is sequential deletion requests. Therefore, for a fixed checkpoint, the noise depends only on the total number of deleted samples, and it grows with this number as in the original CKDU. Extensions of our framework to other sequential unlearning processes, such as by starting from a noisy model, are beyond the scope of our current work.

We also evaluate CKDU under sequential deletion requests on MNIST with a strongly convex student. We consider cumulative forget sets corresponding to 5%, 10%, and 20% of the student training data, where each larger forget set includes all samples deleted in the preceding requests. For each request, the student starts from the same fixed checkpoint (TC=16T_C=16 out of TS=32T_S=32 training steps) and performs the remaining K=16K=16 distillation steps on the updated retained dataset. This allows CKDU to handle accumulated deletion requests without retraining the student from scratch or accumulating noise from previous releases. The results are reported below.

Methodσ\sigmaTestForgetRetainMIA
Total deleted: 5.0%
Original-84.798.0100.066.3
Retrain-84.686.0100.051.4
Redistill-only-84.694.0100.068.6
Retrain + noise0.9083.683.099.650.6
CKDU0.9083.684.099.450.2
Total deleted: 10.0%
Original-84.798.0100.063.9
Retrain-83.785.0100.051.9
Redistill-only-84.693.0100.064.7
Retrain + noise1.7482.082.098.050.9
CKDU1.7481.982.598.150.3
Total deleted: 20.0%
Original-84.9100.0100.063.3
Retrain-84.787.0100.053.1
Redistill-only-84.895.0100.064.1
Retrain + noise3.3674.973.094.650.1
CKDU3.3675.275.495.250.0

While we are gaining efficiency by warm-starting from a checkpoint, the noise level is slightly increasing with the total number of deleted samples aligning with our theoretical analysis. However, even with the increased noise level, the warm-start CKDU is able to achieve comparable utilities while saving half of the unlearning time.

The detailed results are provided in Appendix I.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Main Contributions and Revisions

We thank the Area Chair and reviewers for their time and constructive feedback. We have substantially revised our manuscript and responded point-by-point to all three reviews. We summarize the main changes below, with further details in our individual responses.

Our main contribution is a certification framework for unlearning from student models distilled using black-box teacher predictions. Unlike conventional single-stage unlearning, in this case the retraining reference must account for both the teacher and student, even though the teacher is inaccessible for retraining. To that end, we use stability arguments to bound the counterfactual teacher-prediction shift, propagate this bound through student optimization, and calibrate Gaussian noise to certify student unlearning (Theorems 5.1–5.5).

  • (Extension to broader applications) We now allow distinct teacher and student training datasets with a common forget set. The student operator does not need access to the teacher-only training samples (Section 3).

  • (Reduced unlearning cost) We introduced a new warm-start certificate for strongly convex students, which allows reusing a saved student checkpoint instead of repeating all student distillation steps (Theorem 5.6). We also evaluated cumulative deletion requests using this approach (Appendices H–I).

  • (New experiments) Our new experiments include more informative teachers with convex/non-convex students, and a labels-only retraining baseline. We report the regularity constants and calibrated noise, and demonstrate the utility benefit of distillation (Section 6, Table 1).

  • (Practical setting and assumptions) We provided real-world examples of student-only deletion scenarios (Appendix A). We also clarified the distinction between black-box parameter access and training metadata required for certification, and discussed emerging regulatory practices for sharing technical documentation with downstream developers.

We have highlighted our changes in blue in the revised version, and retained our earlier experiments in the appendix. We appreciate your consideration of our revised manuscript and our detailed responses to the individual reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.