acceptodds
Under review as a conference paper at ICLR 2027

The Shape of Forgetting: Modeling Retrained Outputs for Machine Unlearning

Abstract

Machine unlearning is commonly evaluated against an oracle model retrained without the data to be forgotten. In this paper, we empirically investigate whether the outputs of such a retrained model on forget samples can be approximated using only the original model and forget data. This is particularly relevant when retain data, i.e., the original data excluding the forget samples, are no longer available. We study recurring patterns in the transformation from original to retrained outputs by decomposing it into four questions: which forget samples change argmax, which class loses probability mass, how much mass is removed, and where it is redistributed. We observe that uncertainty predicts argmax-change propensity, removed mass concentrates on the original top-1 class in high-accuracy regimes, gained mass favors high-ranked alternatives, and even argmax-stable samples require calibration. Based on these observations, we construct synthetic retrained targets and distill an unlearned model directly on forget data. An extensive set of experiments on Food-101, Oxford Pets, Stanford Cars, CIFAR-100, Tiny-ImageNet with ResNet-18, DeiT-T, and ViT-S evaluate the approach in forget-only regimes with different fractions of removed samples. We show that an unlearned model distilled from the proposed synthetic targets achieves a favorable tradeoff between forgetting, utility, privacy, and distributional alignment: across random 10% and 20% forgetting settings, it tracks the retrained model more closely than competing forget-only baselines in KL/TV on forget samples, remains competitive in test accuracy and membership-inference evaluation, and is substantially faster than retraining.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
machine unlearning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_kJ7COpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how predictions on forget samples change after retraining without those samples, aiming to approximate these outputs without access to retain data. Its empirical analysis examines which samples change predicted class, how probability mass is redistributed, and how confidence changes when predictions remain stable. Based on these observations, the authors propose Rank-Spread Flip (RSF), which uses entropy-based sample selection, mild confidence calibration, and minimal top-class probability adjustments to construct synthetic targets for forget-only distillation. Experiments across five image-classification datasets with 10% and 20% random deletion evaluate alignment with retrained outputs, forgetting, retained utility, membership inference, and runtime. The reported results generally show improved trade-offs over the evaluated baselines, although substantial utility gaps remain in some settings.

Critical strengths

Useful empirical analysis of retrained outputs. The paper examines which forget samples change predictions and how their probability distributions shift, while accounting for variability across retraining seeds. This provides useful guidance for designing unlearning targets beyond simply forcing misclassification or reducing confidence uniformly.

Promising results under forget-only updates. Across five image-classification datasets and two deletion fractions, RSF generally achieves a better forgetting–utility trade-off than the evaluated baselines. These results support the practical value of selectively adjusting soft targets, although gaps to retraining remain.

Critical weaknesses

The methodological novelty is overstated. Approximating retraining without the forget data is a standard objective of machine unlearning, and forget-only distillation with modified targets already exists in methods such as UNDIAL. Moreover, ignoring the numerical margin, RSF’s changing branch is equivalent to lowering the highest logit to the runner-up logit while leaving the others unchanged. The contribution therefore lies primarily in the empirical observations, sample selection, and adjustment rules, rather than a new formulation or redistribution mechanism. However, the target-level ablation uses an oracle change mask, and matched end-to-end comparisons do not sufficiently isolate the benefit of these specific design choices.

The evaluation does not adequately distinguish output adjustment from removal of sample-specific information. Prior work has demonstrated that classifier changes can produce apparently successful unlearning while leaving recoverable information in internal representations. RSF directly optimizes modified output targets; its distributional comparisons and reported membership-inference results are useful, but do not rule out this failure mode. Since the task is random-sample deletion, the relevant concern is residual sample-specific information relative to retraining—not whether class information remains decodable. Without complementary representation-based privacy or recovery evaluations against retrained controls, the evidence supports improved output matching more clearly than removal of training-data influence.

Questions for authors

Clarification of the methodological contribution: Is the changing branch equivalent, up to the numerical margin, to lowering the top logit to the runner-up logit? Under matched optimization settings and without oracle change masks, what evidence isolates the gains from entropy-based selection over random selection and simpler fixed-strength logit adjustments?

Evidence addressing residual information: Do you have evidence that RSF reduces sample-specific information beyond changing the model’s readout—for example, representation-based membership inference or controlled recovery tests compared with retrained models? Class-label probing alone would not resolve this concern in the random-deletion setting.

Clarification of model selection: What exact criterion selects ρ\rho and the reported checkpoints? Does it use any retrained-model information, and what data would the selection procedure require in deployment?

2 comments

Thank you for the update. Overall I think this is a good paper and should be accepted. I have therefore increased my score accordingly.

AuthorsOpenReview ↗

Response to Reviewer kJ7C

Thank you for your time and for the constructive feedback throughout the discussion.

Reviewer_pcbPOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to approximate retrained-model outputs without access to retain data. The authors analyze prediction changes after retraining and propose Rank-Spread Flip (RSF), which constructs synthetic targets for forget-only distillation. Experiments show that RSF generally improves distributional alignment with retraining while maintaining competitive utility and privacy performance.

Critical strengths
  1. The paper offers an interesting perspective on unlearning by approximating retrained outputs rather than directly suppressing predictions.

  2. The empirical analysis provides useful insights into how prediction distributions change after retraining.

  3. The experiments cover multiple datasets, architectures, and forgetting ratios, with comprehensive evaluation metrics.

Critical weaknesses
  1. Incomplete baseline comparisons: recent forget-only methods such as MU-MIs and pH-Strips are discussed but not evaluated, limiting the comparison with closely related approaches.

  2. Unclear hyperparameter selection:although the paper analyzes sensitivity to the change budget ρ\rho, it remains unclear how to select this parameter in a strict forget-only setting.

  3. Limited component ablations: Appendix D evaluates synthetic targets using an oracle change mask, but does not show how each component contributes to the final unlearned model's performance.

Questions for authors
  1. How does RSF compare with recent forget-only methods such as MU-MIs and pH-Strips, which are discussed but not included in the experiments?

  2. How is the change budget ρ\rho selected in practice without access to retain data or retrained-model outputs?

  3. Can the authors provide end-to-end ablations showing the contribution of each RSF component without using an oracle change mask?

Reviewer_Zcp5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

The paper sets out to investigate how machine unlearning can be carried out when there is no longer access to retain data. The authors consider the classification setting and make an analysis on how unlearning removes mass from the posterior class probability of the top 1 class. They then analyze to which classes this mass is transferred. Based on their analysis, they propose a protocol on how to build new targets for the forget set and learn those with knowledge distillation.

The authors consider the class-balanced random deletion protocol that is commonly used, and test their approach on CIFAR-100, Tiny-ImageNet, Food-101, Oxford Pets, and Stanford Cars datasets. The method is compared to 8 other machine unlearning methods, 4 which are specifically designed for unlearning without retain data, and 4 that are not. For the latter 4, the then disable the "retrain" step. Overall their methods works well and demonstrates best alignment with the retrained model on most cases.

Critical strengths
  1. The proposed method works without the retain data - this makes the method practically applicable in cases where you either have large datasets or are not allowed to store the data after training.
  2. The methods show very good performance compared to the baselines, and the code is released which makes the experiments easier to reproduce.
  3. I like the way that their method is motivated by understanding how probability mass is redistributed and then mimicking that. It's not an idea that I've seen elsewhere (but I haven't looked exhaustively.)
Critical weaknesses

While I do understand the overall idea of the paper, I have difficulties in understanding it fully. This is largely due to convoluted sentences and terminology and variables are not properly introduced. Granted, some of these are implied and well-known (e.g KL, TV, F_1 are not defined, they should be), but it should always be defined clearly what is meant by symbols.

Some of the examples of symbols that are not defined (please note, this is not at exhaustive list of all undefined symbols): L140: \Delta_^{C-1}, L141: mathcal{I}_i, L158: S is not defined, neither is p_i^{R_s}

Font size in Table 1 is ok (font size approx 6.5) but in Table 4 it is way too small to read on a print out (approx font size 4). I'll leave it to the AC to assess if this is a problem according to the template, and a potential break of the formatting instructions. Font sizes of Figure 1,2 and 3 are also problematic and should be increase

Questions for authors
  1. Since I have a hard time understanding the paper in details, in order to provide a better assessment during the rebuttal, please provide a clear description of all abbreviations and symbols.
  2. Please number all equations, it makes it easier to refer to equations
  3. L167: How is "top1 share" in table 2 exactly computed?
  4. L206: the claim "is to lose the argmax, probability must leave it" is only true of the probability is above 0.5. Otherwise a class may loose it's top rank by mass being collected by other classes, e.g consider the case p_max = 0.4 and p_2 = 0.39 and p_3 = 0.21. By transferring mass from p_3 to p_2, p_2 becomes the new arg max while p_1 is unchanged. How does this fact change your analysis?
  5. Figure 1, you use the word "quintile": you mean quantile? I'm a bit unfamiliar with "quintile" in this context
  6. You use the symmetric KL, did you consider the Jensen Shannon Divergence that is often used in MU to compare posterior probabilities?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of revisions

We thank the reviewers and the AC. New content is marked in dark red in the updated version of the paper. The main changes:

  • Presentation (Reviewer Zcp5). We reduced and clarified the symbols and concepts introduced in the main text, and added a glossary of notation (Appendix A). We numbered all equations used in the main text. We also simplified the opening of Section 5 and the paragraph “How much mass should move?” in Section 4, and reorganized the appendix (new Appendices J and K).
  • Scope of the claims (Reviewer kJ7C). The paper now presents RSF as constructing synthetic targets that approximate the retrained model’s outputs on forget samples. The conclusion claims closer output matching than the evaluated forget-only baselines, with competitive utility, and the Limitations (Section 7) state that RSF does not constrain internal representations.
  • Additional baselines (Reviewer pcbP). MUpHT (pH-Strips) and MU-MIs are added to the main results.
  • Additional metric (Reviewer Zcp5). We added the Jensen–Shannon divergence to the outputs of the retrained model, defined in Appendix B, to the full results in Table 4 and Appendix I (Tables 18–20).
  • End-to-end ablation (Reviewer pcbP, Reviewer kJ7C). Section 6 and Table 6 replace one component at a time: no changing branch, no gate, random gate, and a fixed instead of adaptive reduction, under the same ρ and optimization settings as the main results.
  • Choosing ρ without the retrained model (Reviewer pcbP, Reviewer kJ7C). Section 6, Table 5 and Appendix J describe a bisection rule that uses the original model’s validation accuracy as a proxy for the retrained forget accuracy, compared with retrained-based selection and forget-only alternatives.
  • White-box audit on internal representation (Reviewer kJ7C). Appendix K reports a representation-based membership inference attack and a controlled recovery test, with the retrained models and Retrain-KD as references.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.