acceptodds
Withdrawn from ICLR 2027

Generalization Guided Machine Unlearning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep machine unlearning has become increasingly critical and urgent across various applications. Existing unlearning methods typically modify the model’s weights to scrub the information associated with the forget set. However, these methods often fail to fulfill the effective and efficient unlearning goal and also expose serious privacy leakage risks for data to be forgotten. Rethinking the machine unlearning problem through the lens of model generalization, we provide a fundamental understanding of the essentials of machine unlearning and validate the privacy risks of deep models trained under various generalization-improving policies. Building upon the intrinsic insights and gained key observations, we propose generalization-guided unlearning methods, GENUN, consisting of GENUN-M by leveraging a high-bias checkpoint model and GENUN-F by employing targeted rewinding and unlearning via soft targets derived from contrastive learning. Extensive empirical evaluations demonstrate that GENUN algorithms consistently achieve near-ideal forget quality and utility while maintaining significant speed-ups. They outperform various state-of-the-art baselines and can protect the privacy of the forget data effectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Data Privacy, Privacy Preservation, Model Generalization

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_AGC7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines machine unlearning (MU) through the lens of model generalization. Based on the assumption that a well-generalized model depends less on individual samples, which makes unlearning easier and reduces membership inference (MIA) risk, this paper proposes two unlearning methods, GENUN-M and GENUN-F. As supporting evidence, they observe how various data augmentation techniques and training policies affect the generalization gap and MIA attack success rate (ASR) for ResNet-18 on CIFAR-10. GENUN-M assumes access to an intermediate checkpoint Θ^k\hat{\Theta}_k with the smallest generalization gap. Starting from this checkpoint, it optimizes a loss commonly adopted by many existing unlearning methods by fine-tuning on Dr\mathcal{D}_r combined with gradient ascent on Df\mathcal{D}_f. GENUN-F targets the standard unlearning setting where only the final model is available. It jointly optimizes supervised learning on Dr\mathcal{D}_r and a MoCo-based InfoNCE loss that pulls the representation of each forget sample toward its augmented version while pushing it away from neighboring samples in Dr\mathcal{D}_r . The methods are compared with 11 baselines on CIFAR-10/100, SVHN, and Tiny-ImageNet, mainly under random-subset forgetting. The evaluation uses accuracy-based metrics, running time, MIA ASR with simple classifier-based attackers, and U-LiRA for part of the experiments.

Critical strengths

The research question “how generalization relates to machine unlearning” is meaningful and worth investigating.

Critical weaknesses
  1. The core hypothesis that “well-generalized models and checkpoints contain little information about Df\mathcal{D}_f” is supported neither theoretically nor empirically, and it contradicts the assumptions of the proposed method. Theorem 1 is labeled informal, but it does not have the form of a theorem at all. In Theorem 2, the uniform stability condition (T1) essentially assumes the conclusion, so the generalization-gap condition plays no role. Under (T2)/(T3), the step leading to the per-sample influence bound is not proven. In addition, the “mild assumption” in Sec. 3.3.1 is never stated, and empirical observations are presented as lemmas (Lemmas 1–2). The observations supporting the assumption (Observations 1–2) rely on a single experimental setting, CIFAR-10/ResNet-18. They use a one-time test accuracy gap as the measure of generalization and do not distinguish between overfitting and memorization, a distinction drawn by recent work. As Feldman (2020) and others have shown, memorization can be necessary for generalization on long-tailed data, but the paper sidesteps this with the undefined notion of a “typical data point.” Moreover, a small average gap does not imply that the model holds little information about individual samples. GENUN-M assumes that Θ^k\hat{\Theta}_k “contained limited sample information by construction.” However, Θ^k\hat{\Theta}_k is part of the training trajectory on D\mathcal{D}, which includes Df\mathcal{D}_f (k=1k=1 in the experiments). By definition, it has already been influenced by Df\mathcal{D}_f. The selection criterion for kk, GAccG_{Acc}, is also computed on D\mathcal{D}, which includes Df\mathcal{D}_f. The paper provides no clear theoretical justification and no empirical evidence that the influence of Df\mathcal{D}_f on Θ^k\hat{\Theta}_k is indeed small. To claim this as a general principle, consistent verification across diverse datasets and models would be required. If this assumption does not hold, the results of GENUN-M may reflect starting from an under-trained state rather than actually forgetting Df\mathcal{D}_f, which undermines the core justification of the proposed method.
  2. The evaluation does not match the goal of instance-level unlearning, and the proposed method (GENUN-F) may in fact make forget samples more distinguishable. In instance-level unlearning, the goal is to verify whether the information of specific examples has actually been removed from the model, not to make the accuracy on Df\mathcal{D}_f similar to the test accuracy. Unlike class-level or concept-level unlearning, a random subset is representative of the data distribution at the instance level, so a well-generalized model behaves similarly to the retrained model to begin with, and accuracy-based metrics barely measure whether forgetting has occurred. Furthermore, the retrained model is a distribution rather than a single model due to training stochasticity, so this needs to be considered. For empirical unlearning without theoretical guarantees, the criterion for “the unlearned model is similar to the retrained model” should be clearly defined, and its relation to certified unlearning should be discussed. The privacy evaluation does not support the claims either. The MIAs used achieve an ASR of at most 0.53 even before unlearning (Fig. 1d), and the MIA (loss) on the Original model is 55.5 (Table 4), so an ASR of about 0.5 cannot be taken as evidence of successful unlearning. In addition, the contrastive objective of GENUN-F (Eq. 3) pushes the representations of forget samples away from nearby retain samples. In the retrained model, however, a typical forget sample is treated as an unseen sample and is expected to be represented close to similar retain samples of the same class. This objective may thus move in the opposite direction of the retrained model’s behavior. Moreover, if all forget samples are systematically pushed away from the retain distribution, they may acquire a shared characteristic that distinguishes them from test samples, increasing privacy risk. Thus, the current evaluation cannot determine whether the proposed methods actually remove the information of the forget samples.
Questions for authors
  1. What evidence supports the premise that Θ^k\hat{\Theta}_k contains “limited sample information”?
  2. Due to training stochasticity, retrained models form a distribution. Could you report the variance of each metric over retrained models with multiple seeds, and show where the unlearned models lie within that distribution?
  3. For the MIAs in Figs. 1, 5, and 8, on which data were the attackers trained, and how were they evaluated? If these attacks are near chance even before unlearning, why can they serve as evidence of successful unlearning?
  4. Several implementation details of GENUN-F are unclear. (1) How is the neighbor set constructed: using representations from which model and which layer, with which distance metric, and how many neighbors? Is this computational cost included in the reported running time? (2) In the positive key k+=Fenc(x~f;Θ)k^+ = F_{enc}(\tilde{x}_f; \Theta), which parameters does Θ\Theta refer to: the original model, the query encoder, or the key encoder? (3) The default value α=5\alpha = 5 in Appendix D contradicts the definition α∈[0,1]\alpha \in [0, 1] in Eq. 2, and with this value the sign of the contrastive term is reversed. Please clarify the actual value and objective used. Please also explain the choice of τ=3\tau = 3.
  5. GENUN-M and GENUN-F appear to combine two independent design choices: initialization (Θ^k\hat{\Theta}_k vs. Θˉ\bar{\Theta}) and the forgetting objective (gradient ascent vs. InfoNCE). The motivation of GENUN-F, that the ideal output on Df\mathcal{D}_f is ambiguous, applies equally to the gradient ascent term in GENUN-M. Why is the contrastive objective not used in GENUN-M?
  6. Can Theorems 1 and 2 be presented with rigorous statements and proofs? If not, it would be more appropriate to recast them as conjectures or discussion.
Reviewer_VXGgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies machine unlearning from the perspective of model generalization and proposes two methods, GenUn-M and GenUn-F. GenUn-M uses an intermediate checkpoint with a smaller generalization gap as initialization, while GenUn-F employs contrastive learning to guide forgetting when only the fully trained model is available. Experiments across several image classification datasets and model architectures compare the methods with existing unlearning approaches in terms of accuracy, efficiency, and membership inference attacks. The reported results are promising, particularly for GenUn-M.

Critical strengths
  1. The paper offers an interesting perspective on machine unlearning by connecting forgetting with model generalization. The distinction between checkpoint-based and fully trained model settings is practically relevant, and the two proposed methods address these settings separately.
  2. The experimental evaluation covers several datasets and model architectures, with comparisons against established unlearning baselines and retraining. The inclusion of membership inference attacks, including U-LiRA, strengthens the empirical evaluation. GenUn-M achieves encouraging results in terms of accuracy and computational efficiency.
Critical weaknesses
  1. The theoretical analysis contains serious correctness issues that undermine the paper's main theoretical contribution. In Appendix B.2, Lemma 5 (Eq. (5)) claims a positive lower bound on the difference between P(hl∣Dr)P(h_l | Dr) and P(hl∣Dr Df)P(h_l | Dr \ Df). However, DrDr and DfDf are disjoint by definition, so DrDr \ DfDf = DrDr and the left-hand side is identically zero. This invalidates a key step used in Theorem 3. Moreover, Theorem 2 does not adequately justify the claimed implication from a small generalization gap to bounded sample-specific information in the model parameters. These issues substantially weaken the theoretical foundation of the proposed approach.
  2. The current evaluation does not sufficiently distinguish actual forgetting from improved generalization. The Acc-Index primarily measures the agreement between forget-set and test-set accuracy, which does not necessarily indicate that the influence of forgotten samples has been removed. For example, in Table 1, the original ResNet-18 on CIFAR-10 achieves a higher Acc-Index than the retrained model despite performing no unlearning. Furthermore, GenUn-F still has a U-LiRA AUC of 0.647, noticeably above the ideal value of 0.5. These results raise concerns about whether the reported improvements support the paper's broad claims of near-ideal forgetting and privacy protection.
Questions for authors
  1. In Eq. (3), the keys are described as representations of samples from DrDr, while the positive key is constructed from an augmented forget sample. Is the positive key included in the denominator of the InfoNCE loss? Please clarify the exact formulation and implementation.
  2. Eqs. (1) and (2) restrict the coefficient α to [0, 1], whereas Appendix D reports a default value of α = 5. Which formulation and parameter range were actually used in the experiments?
Reviewer_ePoQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the deep machine unlearning problem via the lens of model generalization. The authors claim that existing unlearning methods may face the challenge of exposing privacy leakage risks for data to be forgotten. The authors also raise concerns about the efficiency of existing state-of-the-art machine unlearning (MU) methods. Based on some insights and observations, the paper proposes generalization-guided unlearning methods GENUN. The paper provides experiments selecting four commonly used datasets for MU.

Critical strengths
  1. The paper introduces an important topic of accurate deep machine unlearning while effectively protecting the privacy of the forgotten data. It presents an interesting investigation into model generalization, privacy, deep machine unlearning, and efficiency, and therefore proposes generalisation-guided algorithms together with theorem analysis.
  2. The paper conducts extensive evaluations of methodologies across 4 commonly used datasets. The results are also positive; the proposed GENUN method outperforms all baseline methods, achieving promising performance across different metrics.
Critical weaknesses
  1. The key research question is unclear. The authors mentioned three issues of machine unlearning, i.e., lack of privacy, accuracy, and efficiency, making the main aim and research problem unclear.

  2. Intuitively, I cannot understand why the proposed solution will effectively address the research problem. Without seeing the experiments, I am not confident that the problem will be very well addressed. The authors should intuitively make it clear why the proposed solution work in the introduction.

  3. The latest baseline was published in 2024. The missing of sufficient latest baselines at top venues makes the experiments and conclusions less convincing.

  4. The datasets are of small scale and the models are limited to vision tasks.

Questions for authors
  1. Can you clearly articulate the research problem?

  2. Can you intuitively explain why the proposed methods work?

  3. Can you add sufficient latest baselines from top venues?

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

Related papers

Discussion (0)

Sign in to comment.