I have critical concerns regarding what the paper actually achieves. It seems to me that the paper creates a superficial problem and analyses it. The paper neither gives a solution to its created problem nor analyses an actual problem of practical importance, making me recommend rejection in its current form. I clarify this position in the comments below.
W1 [CRITICAL]. The paper assumes access to only partial information without clear specification of the practicality of that assumption in practice:
- i) The authors cite papers proposing unlearning methods under partial information (e.g., Guo et al. 2020, Golatkar et al. 2020, Izzo et al. 2021). However, those cited works do not restrict unlearning to a partial-information style, they simply use partial information to induce unlearning from a practical point of view. In other words, they argue that the partial information they use suffices to induce unlearning. Citing them to motivate a new formulation of unlearning under partial information as if it is the norm in unlearning practices inflates the problem. What would be an interesting highlight is to investigate the limits those methods can best achieve in approximating the retrained model in practical experimental settings.
- ii) The paper does not tell us what is the required partial information to remove the ambiguities of interest in unlearning. One doesn't have to have complete information about the data generating distribution for example to be able to perform perfect unlearning. Having partial information is fine, but what is the sufficient partial information to alleviate their raised bounds? Discussing and explaining this is fundamental to this work, and needed. I understand that this is stated as a future direction in the conclusion, but I see the work without that discussion quite insignificant.
- iii) In practice, the unlearner has access to the full training dataset of retained and to-forget data of both features and labels. Is this considered partial information? This is the most practical case and it is not discussed in this work.
- iv) The definition of partial information given in this work is a set of real-valued scalar functions. It is not clear how that encompasses the case when the available partial information is the data without labels or the loss function local curvature. I think the definition of the class of functions F=f:D→R requires a revisit to fit the types of partial information considered in the paper.
- v) Section 3 only deals with ambiguity related to lack of labels. As I pointed above, in practice the labels are almost always available. The reader would benefit a lot from practical unlearning scenarios in which this limitation holds. Moreover, what about other partial information cases? For example, what about the lack of one or multiple features of the data while having access to the labels? What about using only a subset of the retained or to-forget data instead of the whole datasets? Or, using a proxy of retained or to-forget data distributed according to the same distribution as those but not used for training (e.g., one is trying to forget all peacock images in CIFAR100 but has access to only peacock images from the test split not the training split)? If the focus is only on special partial information such as lack of labels or accessing only local second-order approximations of the retained loss (as is the case in sections 4 and 5), then that has to appear in the title as well as the abstract of this work. The work currently does not reflect a general treatment of all partial information types.
- vi) The retained and forget risks are unjustifiably of cumulative loss. Why are not they averaged over the number of data points which is the practical formulation? In that case, the limits in Theorem 1 would not depend on the size of the forget set, deeming some of the central arguments of the work lapse.
W2. This work focuses on approximating the retrained model in terms of its outputs. This is stated in lines 302-303.
Intuitively, the overall target of unlearning, approximating the retrained model,
can be interpreted in two ways: (i) matching the empirical loss of the retrained model on the retained
dataset; (ii) matching the prediction of the retrained model on the forgetting data to prevent privacy
leakage.
However, this is not the most common definition of approximate unlearning in the literature. Researchers are interested in approximating the retrained model in the parameter space rather than the output space. This seems to be implicitly targeted in the Toy Non-Linear Example starting in line 340 (particularly the equations in line 356), but it is deeply vague and confusing what the authors mean by unlearning (referred to at times as "exact", e.g., in line 620). For this matter, it is important to note that it is uncommon in the unlearning literature to target one specific retrained model, as that is an infeasible target in practice. The common target is statistical indistinguishability in the parameter space which accommodates stochasticity in what the unlearning algorithm actually achieves (cf. Guo et al. 2020). In other words, not having exactly one retrained model for a given unlearning problem is technically fine as far as the indistinguishability guarantees are concerned. This requires careful treatment from the authors in order to pinpoint the actual scale of the problem induced by the ambiguity in the dataset given the partial information.
Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal
from machine learning models. In International Conference on Machine Learning, pp. 3832–3842.
PMLR, 2020.
Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net:
Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer
vision and Pattern Recognition, pp. 9304–9312, 2020.
Zachary Izzo, Mary Anne Smart, Kamalika Chaudhuri, and James Zou. Approximate data deletion
from machine learning models. In International conference on artificial intelligence and statistics,
pp. 2008–2016. PMLR, 2021.
Response to Reviewer yQW5
We thank Reviewer yQW5 for the positive assessment and helpful comments. We respond to the comments below.
Q1 and W2: Extension to binary classification with cross-entropy loss
We thank the reviewer for this insightful question. To address this question, we have added Appendix G, which extends the retain and forget risks decompositions to binary classification with cross-entropy loss. Specifically, Propositions 6 and 7 express the risks through algorithm-dependent KL fitting terms and algorithm-independent entropy terms, analogous to Propositions 1 and 2. The fitting terms measure agreement with the conditional mean labels and the conditional mean retrained predictions, respectively. Their non-negativity also yields lower-bound inequalities analogous to Corollary 1.
Q2 and W1: Extension to real-data unlearning scenarios
We agree with the reviewer that empirically assessing the practical implications of our results on real-world datasets is an important direction for future work. Our intuition is that similar information limitations may also hinder unlearning in real-data settings. As illustrated in Figure 1, the difficulty arises from the unlearner’s limited information about the retained-data loss and its curvature outside a local neighborhood of the original model. In practice, the high dimensionality of the parameter space and the complexity of real-world data distributions may further complicate inferring the loss landscape from these local observations.