- 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 with the smallest generalization gap. Starting from this checkpoint, it optimizes a loss commonly adopted by many existing unlearning methods by fine-tuning on combined with gradient ascent on . GENUN-F targets the standard unlearning setting where only the final model is available. It jointly optimizes supervised learning on 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 . 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
- The core hypothesis that “well-generalized models and checkpoints contain little information about ” 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 “contained limited sample information by construction.” However, is part of the training trajectory on , which includes ( in the experiments). By definition, it has already been influenced by . The selection criterion for , , is also computed on , which includes . The paper provides no clear theoretical justification and no empirical evidence that the influence of on 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 , which undermines the core justification of the proposed method.
- 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 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
- What evidence supports the premise that contains “limited sample information”?
- 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?
- 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?
- 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 , which parameters does refer to: the original model, the query encoder, or the key encoder? (3) The default value in Appendix D contradicts the definition 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 .
- GENUN-M and GENUN-F appear to combine two independent design choices: initialization ( vs. ) and the forgetting objective (gradient ascent vs. InfoNCE). The motivation of GENUN-F, that the ideal output on is ambiguous, applies equally to the gradient ascent term in GENUN-M. Why is the contrastive objective not used in GENUN-M?
- 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.