- 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
- 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.
- The theoretical chain is modular and appears correct, and the convex values in Table 4 can be reproduced from Eq. 10.
Critical weaknesses
- 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.
- 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
- How can Table 1 show a 19-point forget-accuracy gap when Fig. 2 bounds the logit difference to about ? An explanation could restore my Soundness score, while a failure of the bound would lower my rating to 1.
- Which strong-convexity constant and teacher step schedule were used to compute the noise in Tables 4 and 7?
- What step size was actually used in App. H, and how far did the teacher's parameters move?
- Can you show a setting with an informative teacher, valid constants and reasonable utility? This would most likely raise my rating.
4 comments
[W2, Q2] Thanks for catching this. We noticed that this was due to a typo, the distillation coefficient should be 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 reported in Table 2 (now Table 10 in the revised version), which is for XOR and for MNIST and CIFAR-10. The teacher is trained by single-sample SGD for steps with the step size on XOR and 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 , that is, over 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 , so over the steps the teacher moves at most from its pretrained initialization, which is at most . However, the regularity constants of a ResNet-56 are large ( 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 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:
-
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.
-
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 on the forget samples and 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 difference between their logits as on the forget samples and 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 , 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.
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 on MNIST and λ=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%.
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)
CIFAR-10 (16D PCA)
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 BJ 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.