acceptodds
Withdrawn from ICLR 2027

Advancing Machine Unlearning Evaluation Requires Rethinking Retraining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine unlearning is emerging as a pivotal paradigm aimed at reducing the impact of unwanted data points previously used in training machine learning models. Despite recent advancements in developing new machine unlearning algorithms, robustly evaluating the effectiveness of these algorithms remains a challenging problem. Retraining a model from scratch, which many algorithms seek to circumvent, continues to be a popular benchmark for comparison due to its inherent nature of excluding undesirable data from the outset. In this study, we critically assess the role of retraining as a benchmark for evaluating machine unlearning algorithms. Through extensive experiments across various evaluation metrics, datasets, and model architectures, we uncover notable instability in retrained models. Our findings reveal that retrained models can exhibit both the highest and lowest effectiveness in unlearning, while maintaining the same utility. Additionally, this phenomenon has been shown to be consistent across multiple membership inference attack and distance function metrics. Although our results do not negate the importance of retraining as an unlearning method with guaranteed data removal, they highlight the need for its careful application in machine unlearning evaluation. Code for reproducing all results will be released upon acceptance.

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

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zvczOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the reliability of using a single retrained model as the reference for evaluating machine unlearning. Through experiments across multiple datasets, architectures, forgetting settings, and evaluation metrics, the authors identify two sources of instability: retraining variance, where models with nearly identical utility obtain different unlearning scores due to random seeds or training configurations, and metric fragility, where some metrics fluctuate even for a fixed model. In particular, MIA-R and distance-based metrics are shown to be unstable, while MIA-F and accuracy-based metrics are comparatively consistent. The paper therefore recommends treating retraining as a distribution represented by several checkpoints, reporting reference intervals instead of point estimates, and prioritizing robust metrics such as MIA-F. Overall, the work highlights an important weakness in current unlearning evaluation practices and proposes practical guidelines for more reliable benchmarking.

Critical strengths

This paper identifies an important yet overlooked issue in machine unlearning evaluation: existing studies often treat a single retrained model as an absolute reference. The authors demonstrate that retrained checkpoints with nearly identical test accuracy can produce substantially different unlearning scores, potentially undermining the reliability of existing evaluations. Extensive experiments across diverse settings provide solid evidence of this problem and its practical impact. In particular, the detailed analysis of MIA-R’s failure modes offers valuable insights and may encourage future work on developing more reliable evaluation protocols and metrics.

Critical weaknesses

The main weakness of this paper is its relatively limited technical contribution, as it primarily analyzes evaluation instability caused by randomness rather than proposing a new unlearning method or metric. Nevertheless, the identified problem is important and meaningful to the field. In addition, the stability of MIA-F does not necessarily demonstrate its effectiveness: although MIA-F remains close to 0.5 for retrained models, the paper does not show that it can reliably distinguish successful, partial, and failed unlearning. Finally, the recommendation to use three to five retrained models is not supported by a convergence or sample-size analysis. A more detailed study of how many checkpoints are needed to obtain stable reference intervals would strengthen the paper’s practical contribution.

Questions for authors
  1. Can the authors evaluate whether MIA-F can reliably distinguish successful unlearning, partial unlearning, and complete failure to unlearn, rather than only showing that it is stable on retrained models?
  2. Can the authors provide a convergence or sample-size analysis to justify why three to five retrained models are sufficient for estimating stable medians, confidence intervals, or interquartile ranges? I would be willing to raise my score if the authors can adequately address these two concerns.
Reviewer_tioBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether retraining can serve as a reliable reference for evaluating approximate unlearning. By examining retrained models across different training configurations/datasets/evaluation metrics, they find that models can exhibit substantially different unlearning metric scores while preserving similar utility. They therefore argue for treating retraining as a distribution rather than relying on a single reference checkpoint.

Critical strengths

The paper studies the reliability of the current unlearning evaluation, which is a relevant question in the field. The experiments are comprehensive and cover multiple datasets/architectures/evaluation metrics. I particularly appreciate the additional analyses that distinguish variability across retrained models from randomness in membership-inference evaluation. The fixed-configuration experiments also provide useful evidence that some metrics can vary across random seeds.

Critical weaknesses
  1. My main concern is the research question itself and the experimental design. The main experiments use various training hyperparameters, but exact unlearning is defined relative to a particular training algorithm, including its training configuration. Changing these hyperparameters effectively changes the reference learning algorithm itself. I'm not sure why we would want to test various training configurations in this context, as doing so introduces additional confounding factors that would not arise when evaluating against a fixed retraining procedure. Therefore, I'm not fully convinced that differences between models trained with different configurations demonstrate instability of retraining as an evaluation reference. Moreover, achieving similar test accuracy does not necessarily imply that these models are equivalent in their behavior or output distributions.

  2. The authors show that MIA-R can exhibit substantial variance even when evaluated on the same retrained model. However, as they acknowledge, this can arise from degenerate membership classifiers predicting all samples as members or nonmembers. This seems to be a limitation of the evaluation metric, instead of retraining itself. Similarly, although MIA-F appears more stable in the experiments, it does not necessarily establish that it's a reliable unlearning metric. I would appreciate a clearer distinction between instability caused by retraining and limitations inherent to the evaluation metrics.

  3. The authors recommend reporting retraining as a distribution rather than a single checkpoint, but if different training configurations are included, the resulting distribution depends on how these configurations are selected and weighted. Can the authors clarify how such a reference distribution should be defined in practice?

Questions for authors

Please see the critical weaknesses above. In particular, I would appreciate clarification on:

  • Why varying the reference training algorithm is appropriate, given that exact unlearning is defined relative to a fixed algorithm?
  • How should the proposed reference distribution of retraining be constructed?
Reviewer_mtHEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper questions the use of a retrain model as the perfect unlearning comparison, as some unlearning evaluations can yield very different results depending on the hyperparameters used for retraining. To show this, the authors perform retrained models, on a wide range of hyperparameters. Lastly it makes a comparison between MIA-F and MIA-R, to show that MIA-F is much more robust to hyperparameter variations. They propose to evaluate two sources of instability, the retraining variance and the metric fragility. The authors provide a wide set of benchmarks and evaluations, over many retrained models

Critical strengths
  • The empirical evaluation is extensive, for the amount of retrained models and datasets.
Critical weaknesses
  • It is not clear whether the points in the plots (eg. Figure 1 and 2) are all finished converge models, or they are a mixture of converged models and mid checkpoints, from where we could expect that the evaluation would yield different results, even at the same test accuracy. The reason for this doubt is that in Section 5 the authors mention that they use retrained checkpoints when aggregating the results. Also, in hyperparameter selection they train three models for each hyperparameter combination. Are all of them in the plots?
  • Models with the same utility (accuracy), but with different configurations, like epochs or weight decay (Table 4) will have different generalization gaps. So different MIA-R values are expected. Table 6 shows that changing the weight decay changes the MIA-R distribution.
  • The issue of retraining is shown on a small scale, where retraining is cheaper. The authors use small vision models, and only a bert-base for text. Although the authors have it in their list of models, ViT results couldn’t be found in the paper. It could have been interesting to see comparison at scale, or for larger models, where the retraining becomes limiting or even prohibitively expensive. The authors claim that they use the most widely used datasets and models, but machine unlearning has been tested and tried in bigger models and datasets. For language models, TOFU (Maini et al, 2024) and MUSE (Shi et al, 2024) (that is used in this work) use much larger models, and for vision models the authors can check LoTUS (Spartalis et al, 2025).
  • The authors missed on having evidence for MIA-F, if it can separate the original model from the retrains. This is because for a retrained model, the forget set and the retain set are both non-members. This weakens the headline recommendation.
  • Only one unlearning method, SimNPO, is evaluated, in the A.5 appendix. There is no ranking to compare. The authors could test the approximate methods that they cite (SCRUB, l1-sparse) on one of the model sizes.

The paper presents gaps in its writing and design choices that make the reviewer question the soundness of the experiments. While the problem that is being studied is relevant, it falls short at making any further contribution, like a more robust evaluation metric than MIA-F, or, as the authors propose, a distribution aware reference (Section 7). The work would benefit from a better description of the experiments, what is included on each plot, and what are the main takeaways from them. Also, from validating how the experiments change conclusions or rankings, from previous baselines or benchmarks.

Questions for authors
  • For Table 1, are the values a range or a set? The hyperparameter selection paragraph says it is a range.
  • Figure 3 has a typo (not full word) and the figures confuse the Row for Column.
  • Figure 1 and 2, some plots miss the original pretrained model. The MIA-F and 1-WD have no retrained points at all?
  • In Appendix A.9 (could be a new: Appendix B) the plots are not clear, because they are a continuous line. Is every value on the line a model? If so, it would have been better as a scatter plot, as we have been presented throughout the paper. Else it seems like the training behavior of one model.
  • In Section 4.2 it is said: "MIA-R (correct prob) predominantly yields perfect scores". But in Figure 1, the points are at 0. Is that a typo? Or a different interpretation from the plots?
  • Why not evaluate more recent MIA metrics, like LiRA/U-LiRA/LiRA-Forget?
  • The authors claim: "In Figure 6, (...) using a fixed retrained model at a time, thereby eliminating retraining variance within each conditioned analysis." And later: "For each MIA setting, we use 10 data splitting and shuffling seeds and 10 parameter initialization seeds, which results in 100 classifiers per configuration. We report the key statistics of these 100 scores in Figure 6". Then, for Figure 6, which model is being used? And what we are presented is aggregated from each time? If aggregated, wouldn't that include retraining variance?
  • Definition 2.2 has a mistake, the second row should be: P[U(A(Dtr),Df,Dtr)∈R]≤eϵP[A(Dr)∈R]+δ\mathcal{P}[U(A(D_{tr}),D_f,D_{tr})\in R] \leq e^{\epsilon}\mathcal{P}[A(D_r)\in R] + \delta

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

Related papers

Discussion (0)

Sign in to comment.