acceptodds
Under review as a conference paper at ICLR 2027

Rethinking the Verification of Machine Unlearning

Abstract

Machine unlearning claims require evidence that a model has removed the influence of specified training data. Yet widely deployable post hoc verifiers infer unlearning largely from predictive behavior, the final model response that remains under the provider's control. We make this limitation concrete through Output-Level Deceptive Response (ODR), an attack that produces responses accepted by output based verifiers while preserving the original representation state. By separating output conformity from internal unlearning, ODR motivates a rethinking of machine unlearning verification, shifting the audit target from what the model returns to what changes inside it. We operationalize this shift through Representation Unlearning Verification (RUV), a gray box verifier that compares intermediate forward representations in the original and unlearned models. RUV traces a four-stage response through activation reorganization, representation displacement, neighborhood disruption, and loss of manifold support. Matched retained controls distinguish target specific responses from global model drift, while joint permutation calibration combines the four evidence views into an adaptive decision. Experiments spanning exact and approximate unlearning, class and sample level requests, multiple datasets, and diverse architectures show mean ODR deception success rates of and , respectively. RUV achieves mean verification accuracies of and , while rejecting and of the corresponding ODR submissions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine unlearning, unlearning verification, representation auditing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4iX1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first shows that output-based verification in machine unlearning can be fooled. The paper presents an attack, output-level deceptive response (ODR) to prove the vulnerability. Specifically, the paper shows ODR in two learning scenarios ODR-Head for class unlearning and ODR-Gate for sample unlearning. The attacks pass output-based verifiers even though the original backbone of the model remains intact. Then, the paper proposes RUV framework based on four views: activation route shift, representation shift, neighborhood persistence gap, and manifold rank drop. The proposed verification framework is gray-box unlike previous black-box methods (vulnerable to ODR) and white-box (requires full access, not practical). The proposed gray-box settings require only forward representation, not parameters, not gradients. Experiments are carried out on CIFAR10/100 and FaceScrub with ResNet models and RUV shows good performance. The paper also presents results ViT-S/16 and MobileNetV2 for applicability in different architectures.

Critical strengths

The paper shows attack and fix together with a more realistic threat model. ODR shows output statistics are weak and covers both class and sample request. Three requirements of RUV (attribution, coverage, and calibration) are useful. The four views follow a coherent story (route, state, neighborhood, manifold). The threat model does not require parameter nor gradient access. Evaluation scope covers exact and approximate unlearning, class and sample requests, several datasets and architectures. Experiments also include ablations for each component.

Critical weaknesses

The claims are a bit broad. The related work says outside machine unlearning verification, there are works on representation analysis. There are works on early to mid of this year inside machine unlearning.

https://arxiv.org/abs/2605.27569

https://arxiv.org/abs/2606.25001

https://arxiv.org/abs/2503.06991

https://arxiv.org/abs/2602.18505

Classifier only fine-tuning (illusion of unleanring, https://arxiv.org/abs/2604.08271) is similar to ODR-Head. Really unlearned (https://arxiv.org/abs/2406.10953) invades MIA-based verification for sample unlearning which is similar to ODR-Gate. Very recent works may be parallel to the paper. But earlier papers may overlap with the paper. In addition, I think it is better to compare the propose method with RULER.

Although the proposed threat model is more realistic, it still assume a trusted extraction channel. I think it may be difficult to comply with the provider because the provide may not want to reveal any part of the model information. The representation may reflect the model information. Experiments are only on small models and datasets. How about ImageNet or large models?

Questions for authors

Will RUV still work for large models and datasets?

What are the limitations of the proposed method?

5 comments
AuthorsOpenReview ↗

Response to Weakness — (1)

Thank you for your interest in and recognition of our work.

Regarding Related Work

We have expanded the discussion of the six studies you mentioned and other relevant work to clarify our contributions. Please refer to Sections 1–2 and Appendix A for details. Our analysis follows.

1. Studies Related to RUV

(1) RULER [1]

Compared with RUV, RULER raises concerns about the reliability of output-level verification and analyzes them using its proposed metrics. However, it does not implement concrete deceptive attacks comparable to ODR.

RULER's M2 already uses retained records for calibration, while M4 constructs an internal reference based on the nearest-neighbor distribution of the retained set. Therefore, RUV's distinction does not simply lie in its use of retained data. Instead, RUV integrates multi-view responses between the original and unlearned models, matched retained-data controls, and request-specific joint decisions that preserve dependencies among the evidence views.

Among RULER's four verification metrics, three (M1, M2, and M3) rely on Retraining, while only M4 does not. However, the authors explicitly acknowledge that M4 is a preliminary attempt and exhibits strong dataset dependence, with degraded verification performance in other settings.

We therefore include retraining-free RULER-M4 as an additional baseline. We follow the datasets, model architectures, class-level and sample-level unlearning requests, and submitted models used in our main experiments. Following the official implementation, we extract representations immediately before the classification head and compute M4. An acceptance interval centered at 0.5 is determined in advance using independent calibration requests. We consistently report acceptance rates of different unlearning methods and rejection rates of ODR.

The results show that M4 has difficulty distinguishing Original from Retraining under certain sample-level requests. Moreover, because ODR introduces no backbone update, its preclassifier M4 statistic matches that of Original, resulting in false acceptance. Under class-level requests, semantic differences between the target and retained classes affect M4's verification results, leading to false rejection of correct Retraining submissions. For example, under CIFAR10 sample-level requests, the acceptance rates of Original and Retraining are 0.876 and 0.883, while the ODR rejection rate is only 0.124; under CIFAR100 class-level requests, the acceptance rate of correct Retraining is only 0.613.

(2) Yong et al. (2026) [2]

This work constructs a reference distribution using independently retrained models to analyze discrepancies between output-level forgetting and representation-level consistency with retraining. It also investigates information recovery through diagnostics such as relearning. Its empirical analysis focuses on class-level unlearning.

In contrast, RUV does not require constructing an ensemble of retrained reference models and supports both class-level and sample-level requests. It makes request-level verification decisions through matched retained-data controls and joint calibration. Although this work also highlights the unreliability of output-level verification, it does not propose a concrete implementation of deception comparable to ODR.

(3) Kim et al. (2026) [3]

This work combines output-level metrics, representational similarity, and downstream transferability in large-scale settings, revealing that some unlearning methods merely modify the classifier or degrade representation quality. It also establishes a unified evaluation framework. However, its evaluation primarily focuses on class-level unlearning and relies on comparisons with retrained models.

In contrast, RUV focuses on post-hoc auditing without requiring an independently retrained reference model, while supporting both class-level and sample-level requests. Although this work also identifies the unreliability of output-level verification, it does not implement concrete deceptive attacks comparable to ODR.

(4) Jang et al. (2026) [4]

This work employs sparse autoencoders to identify class-related features and performs inference-time feature interventions to examine whether target knowledge can be recovered, thereby distinguishing representational suppression from information removal. Although its restoration-based analysis does not require retraining, it primarily addresses class-level unlearning.

In contrast, RUV supports both class-level and sample-level requests without training sparse autoencoders or performing feature interventions. Instead, it audits forward representational responses of the original and submitted models. Although this work also raises concerns about the reliability of output-level verification, it does not implement concrete deceptive attacks comparable to ODR.

AuthorsOpenReview ↗

Response to Weakness — (2)

2. Studies Related to ODR

(1) Gao et al. (2026) [5]

This work primarily investigates class-level unlearning. Through feature–classifier mismatch analysis and classifier-only fine-tuning, it reveals discrepancies between output behavior and retained internal discriminative information.

This construction is directly related to ODR-Head, as explicitly acknowledged in our revision. Our work further examines two manipulation strategies: classification-head modification (ODR-Head) and service-response replacement (ODR-Gate). Together, they cover class-level and sample-level requests and evaluate multiple output-level verifiers without representation-level parameter updates.

(2) Xu et al. (2026) [6]

This work demonstrates bypasses of MIA-based and backdoor-based verification through fine-tuning and proposes a verifier based on influential sample pairs. Its MIA bypass makes target outputs resemble those of same-class nonmembers, sharing ODR-Gate's motivation.

However, Xu et al. achieve output alignment through parameter fine-tuning, whereas ODR-Gate directly returns same-class nonmember posteriors for registered target queries without updating model parameters.

Combined with class-level ODR-Head, ODR-Gate enables a unified investigation of output-verifier manipulability across both unlearning request types.

[1] Georgina Cosma and Axel Finke. RULER: Representation-Level Verification of Machine Unlearning. arXiv preprint arXiv:2605.27569, 2026. DOI.

[2] Teresa Pui Yee Yong, Win Kent Ong, and Chee Seng Chan. Erased, but Not Gone: Output Forgetting Is Not True Forgetting. arXiv preprint arXiv:2606.25001, 2026. DOI.

[3] Yongwoo Kim, Sungmin Cha, and Donghyun Kim. Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols. Engineering Applications of Artificial Intelligence, 167:113785, 2026. DOI.

[4] Yurim Jang, Jaeung Lee, Dohyun Kim, Jaemin Jo, and Simon S. Woo. Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearning. Proceedings of the ACM Web Conference 2026, pages 8505–8508, 2026. DOI.

[5] Yichen Gao, Altay Unal, Akshay Rangamani, and Zhihui Zhu. An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. Proceedings of the 29th International Conference on Artificial Intelligence and Statistics, PMLR 300:4114–4122, 2026. PMLR.

[6] Heng Xu, Tianqing Zhu, Lefeng Zhang, and Wanlei Zhou. Really Unlearned? Verifying Machine Unlearning via Influential Sample Pairs. IEEE Transactions on Dependable and Secure Computing, 23(1):1671–1686, 2026. DOI.

Regarding the Threat Model

Thank you for raising the practical deployment concerns. We clarify the following points.

(1). Trusted representation extraction and the auditing setting.

Our original submission explicitly assumes that the extraction interface faithfully returns intermediate representations produced by the submitted model. This is a fundamental assumption of RUV.

We consider controlled audits where authorized auditors have limited forward access. Providers need not publicly expose representations, parameters, gradients, or training logs to ordinary users.

Services offering no internal representation access fall outside this setting, leaving verification dependent on output-level evidence, whose unreliability our work demonstrates.

(2). Relationship to the access assumptions of existing representation-based verification methods.

RUV follows the fundamental assumptions of existing representation-based verification methods while requiring more restricted access. Unlike white-box methods requiring model parameters, detailed architectures, or backward computations, RUV audits through a forward representation interface. With trusted extraction, it provides internal verification evidence with fewer access privileges and without independently retrained reference models.

(3). Privacy risks and practical deployment limitations.

We acknowledge that gray-box access may introduce privacy risks through exposure of representations and model information. Our auditing setting restricts access to authorized parties rather than a public representation interface. Practical deployment therefore requires safeguards such as access control.

Furthermore, extraction-channel authentication, consistency between audited and deployed models, and protection of representation privacy are complementary requirements beyond RUV's statistical decision mechanism.

These assumptions and limitations are consolidated in Section 7 (Discussion) of the revised manuscript.

AuthorsOpenReview ↗

Response to Q1 and Q2

Response to Q1:

Thank you for your suggestion. We agree that dataset and model scale can affect the generalizability of our conclusions.

To further evaluate the scalability of RUV, we have conducted additional experiments on the ImageNet-1K dataset with the ResNet-152 architecture, using Retraining and SCRUB to examine exact retraining and approximate unlearning, respectively. The experimental protocol is provided in Appendix E.6, with detailed results presented in Appendix G.4 and Tables 12–13.

The experimental results show that RUV maintains good verification performance in this larger-scale image classification setting, providing further empirical evidence of its applicability to larger datasets and deeper convolutional models. Specifically, RUV accepts Retraining at 0.989/0.952 and SCRUB at 0.948/0.918 for class-level/sample-level requests; SCRUB's forgotten-class accuracy and sample-level LiRA attack accuracy are 0.011 and 0.506, respectively.

Response to Q2:

Thank you for your suggestion. We agree that explicitly acknowledging the limitations of our method helps clarify its applicability boundaries and guide future research. Therefore, we have added a new Discussion subsection in Section 7 of the revised manuscript to discuss the limitations of our method and potential directions for future work. The main points are summarized as follows.

1. Application Scenarios. RUV is currently designed and evaluated primarily for image classification tasks and cannot be directly generalized to generative settings, particularly those involving large language models (LLMs). Such tasks differ in their output formats, definitions of target information, and construction of reference data. Further research is needed to investigate appropriate evidence views and calibration strategies for these settings.

2. Privacy Risks of Representation Access. Internal representations provide richer evidence for unlearning verification, but they also introduce potential information leakage risks, such as being exploited for model inversion or data reconstruction. Our work adopts a restricted-access setting intended for authorized auditors. However, this arrangement does not itself provide a privacy guarantee. Developing internal auditing approaches that integrate access control and privacy-preserving mechanisms remains an important direction for future research.

3. Trusted Representation Extraction Assumption. RUV assumes that the representation extraction interface faithfully returns the actual intermediate representations produced by the submitted model. Services that provide no internal access do not satisfy this assumption. Moreover, authentication of the extraction channel and consistency between the audited model and the model actually deployed require additional supporting mechanisms.

4. Detectability and Decision Boundaries. When the influence of target records is small, or when legitimate unlearning does not alter the audited representations, RUV may fail to obtain sufficient response evidence. Furthermore, its decisions are based on the statistical responses of the target group and therefore cannot confirm that every individual record in the unlearning request has been successfully unlearned. Improving detection capabilities for small-scale unlearning requests and developing more fine-grained auditing methods are important directions for future investigation.

Reviewer_8hd4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper studies whether a model provider can satisfy machine-unlearning verification without removing the requested data's influence. It introduces Output-Level Deceptive Response (ODR): modifying only the classifier for class requests, or substituting responses to registered sample queries, while preserving the original internal representations. It then proposes Representation Unlearning Verification (RUV), a forward-only gray-box audit of the original and submitted models. RUV combines activation, representation, neighborhood, and manifold-support statistics with matched retained controls and request-specific joint permutation calibration. Experiments cover class- and sample-level requests, several unlearning algorithms, multiple architectures, and selected non-unlearning controls.

Critical strengths
  1. The paper addresses an important auditing problem. A provider-controlled prediction interface is an insufficient basis for inferring what happened inside a model, and moving the audit to internal computation is a worthwhile direction. The explicit ODR constructions make this concern concrete. I was initially interested in the possibility that an inexpensive forward-representation audit could supply useful evidence even without being a complete certificate of unlearning.

  2. RUV is a thoughtful attempt to go beyond a raw feature-distance test. Retained controls address ordinary model drift, and joint permutation calibration accommodates dependence among the four evidence views. The authors also test benign drift, wrong-target unlearning, and correct-label fine-tuning on the requested target set (Table 3). These controls are more informative than comparing only genuine unlearning with an unchanged model. The component and aggregation ablations help explain the proposed design.

Critical weaknesses
  1. The link between the detected event and successful unlearning remains insufficiently established, including under a weaker interpretation as a necessary-condition check.

When I first read the method, my immediate concern was the sufficiency direction: a substantial, target-specific representation change does not establish that the requested information has been removed. A provider could change how information is encoded while retaining it. A procedure that first unlearns and then relearns the requested data is also a relevant negative control: the final model may differ from the original while again containing the target information. I do not claim that this particular procedure necessarily passes RUV; that requires testing against the actual retained-control and joint-decision rule. The concern is that the audited event and the claimed unlearning outcome are different properties.

I nevertheless continued reading with a more favorable, weaker interpretation in mind. Even if passing RUV cannot certify unlearning, perhaps failing the test reliably identifies a lack of unlearning, or the test provides a useful empirical screening signal. Either could be a contribution. By the end of the paper, however, I was also unconvinced that this weaker role had been established. Successful unlearning is not shown to imply detectable changes in at least two selected views. For example, if the prescribed training procedure learns only a classifier on fixed features, exact retained-data retraining can leave the audited feature extractor unchanged. More generally, a record can have a small effect on the selected representations, and a legitimate deletion need not cross the finite-sample detection threshold. Any necessity claim therefore needs explicit restrictions and signal-strength assumptions.

I appreciate that Section 3 operationally defines RUV in terms of corroborated, target-attributed responses. The difficulty is the subsequent interpretation of that event as verification of unlearning. Theorem D.6 controls rejection of an exchangeable no-response null; this is not a false-acceptance guarantee over all models that retain target information. Theorem D.8 assumes positive effects in at least two views and establishes their detectability; it does not derive those effects from successful unlearning. Likewise, Appendix C.4 analyzes the cost of a reference-guided forgery strategy, without excluding other information-preserving internal modifications. Figure 2 reinforces the distinction between response and completion: executing only 75% of a request still yields acceptance rates of 0.842/0.647 for class/sample requests.

I am not asking for a universal certificate or resistance to every possible attack. My concern is that neither a broadly valid necessary-condition interpretation nor a sufficiently validated narrower screening role is established. The paper should specify its intended guarantee and demonstrate what additional errors its detector can reliably identify.

  1. The closest prior work is not positioned sufficiently accurately, making the incremental contribution difficult to assess.

Jeon et al., "An Information Theoretic Evaluation Metric for Strong Unlearning" (AAAI 2026) [1], already demonstrate that a frozen encoder with a modified head can perform well on black-box unlearning metrics, and introduce IDI to evaluate residual information in intermediate features. Their head-distillation example also uses modified original-model outputs rather than requiring a retrained teacher. The present manuscript cites this work on pp. 2 and 15 among output-based methods and then characterizes those methods as observing only the final returned response. This characterization does not describe IDI, and the experiments do not include an IDI comparison. A direct account of what ODR adds beyond head distillation, and what RUV adds beyond existing internal-information evaluation, is important to judging novelty and significance.

The paper does discuss and experimentally compare TruVRF [2], which is a useful distinction based on access and backward-computation requirements. It also cites the fragility of verification under adaptive behavior [3], although the relationship to the proposed attack model could be developed further. Overall, I see a potentially useful combination of internal evidence, retained controls, and calibration, but the current comparison does not establish how much new auditing capability this combination provides.

[1] Jeon et al. An Information Theoretic Evaluation Metric for Strong Unlearning. AAAI 2026, pp. 22173-22181. https://ojs.aaai.org/index.php/AAAI/article/view/39373

Extended version, including Head Distillation details: https://arxiv.org/html/2405.17878

[2] Zhou et al. TruVRF: Toward Triple-Granularity Verification on Machine Unlearning. IEEE Transactions on Information Forensics and Security, 2025. https://doi.org/10.1109/TIFS.2025.3565991

[3] Zhang et al. Verification of Machine Unlearning is Fragile. ICML 2024. https://proceedings.mlr.press/v235/zhang24h.html

Questions for authors
  1. What precise property should a pass or failure establish: a target-specific representation response, a necessary consequence of successful unlearning within a specified model/algorithm class, or an empirical indicator of deletion quality? If the intended role includes a necessary-condition check, what assumptions connect correct unlearning to detectable effects in at least two views? Please explain how the conclusions should be interpreted when legitimate retained-data retraining produces weak or unchanged audited features, and when incomplete requests pass.

  2. Could the authors add a bounded diagnostic in an existing experimental setting, such as CIFAR-10/ResNet18, using an information-preserving internal update or an unlearn-then-relearn procedure? The goal would be to establish both whether target information remains and whether RUV accepts, while controlling retained utility and keeping the attack independent of the particular audit draw. A fixed-budget probe or head-retraining comparison against retained-data retraining could provide an independent recovery diagnostic, rather than defining deletion success from the execution of a configured algorithm. The paper's existing Table 3 controls are valuable, but do not resolve this adaptive setting. A clear failure boundary would also help establish a useful, narrower scope.

  3. Please correct and expand the discussion of Jeon et al. [1], distinguishing their head-distillation construction and intermediate-feature IDI from output-only evaluation. In a shared setting, what detection capability or cost advantage does RUV add relative to IDI and a simple representation-distance baseline? Please account for differences in access, extra training or retraining, and calibration rather than relying only on aggregate verification accuracy.

Additional related-work context, not a requirement for acceptance: RULER [4] is a directly relevant preprint on representation-level verification. Its M2 uses a retained-record calibration baseline, and M4 uses nearest-retain-neighbor similarity ranked against a retained distribution without a retraining oracle. A discussion distinguishing RUV's multi-view, before/after, request-specific decision from these diagnostics would help position the work. I do not base my recommendation on the absence of a comparison to an unpublished preprint.

[1] Jeon et al. An Information Theoretic Evaluation Metric for Strong Unlearning. AAAI 2026, pp. 22173-22181. https://ojs.aaai.org/index.php/AAAI/article/view/39373

Extended version, including Head Distillation details: https://arxiv.org/html/2405.17878

[2] Zhou et al. TruVRF: Toward Triple-Granularity Verification on Machine Unlearning. IEEE Transactions on Information Forensics and Security, 2025. https://doi.org/10.1109/TIFS.2025.3565991

[3] Zhang et al. Verification of Machine Unlearning is Fragile. ICML 2024. https://proceedings.mlr.press/v235/zhang24h.html

[4] Cosma and Finke. RULER: Representation-Level Verification of Machine Unlearning. arXiv preprint, first posted May 2026. https://arxiv.org/abs/2605.27569

14 comments
AuthorsOpenReview ↗

Response to W1— (1)

Thank you for your interest in our work and for your insightful comments.

Regarding your concern about sufficiency, you point out that the connection between significant, target-specific representational changes and unlearning has not been sufficiently established, raising the concern that the audited event and the claimed unlearning outcome may be fundamentally different in nature. We respond in three parts.

  1. RUV's evidence selection is motivated by empirical observations of the unlearning process. We first observed that unlearning operations induce responses along the representation propagation chain, with different unlearning methods affecting different stages to varying degrees. Based on these observations, we constructed RUV from four complementary evidence views: activation routes, feature states, neighbor identities, and manifold support, to investigate whether these responses can serve as effective evidence for unlearning verification. Therefore, these evidence views have an empirical basis grounded in their relevance to the unlearning process.
  2. RUV imposes additional constraints on acceptable responses. Rather than making decisions based on the absolute magnitude of representational changes, RUV compares the response differences between target samples and matched retained samples, and requires at least two complementary evidence views to jointly satisfy the calibrated decision criteria. The calibration process preserves dependencies among the evidence views, avoiding the simplistic accumulation of correlated responses as independent evidence. These designs aim to constrain common drift, single-view anomalies, and random fluctuations, thereby reducing the risk of incorrectly accepting other model updates as unlearning.
  3. We further conducted controlled experiments to examine whether RUV can distinguish unlearning from information-preserving updates. Table 3 of the original submission already included controls involving general drift, wrong-target unlearning, and fine-tuning on the requested target set using correct labels. The last update changes target representations while preserving strong target prediction capability, yet RUV achieves class-level and sample-level rejection rates of 0.968 and 0.929, respectively. This demonstrates that, under these controlled updates, changes in target representations alone are insufficient to pass RUV. For completeness, RUV rejects benign training drift and wrong-target unlearning with class-level/sample-level rates of 0.989/0.947 and 0.976/0.937, respectively.

In addition, following your suggestion, we have completed a new control experiment involving unlearning followed by relearning. The new results are presented in Section 6.3 and Table 3 of the revised manuscript, while the complete relearning trajectories and independent recovery diagnostics are provided in Appendix G.3 and Tables 10–11. The experimental setup and further analysis are detailed in our subsequent response to Q2. Starting from a model that has just completed Retraining, we evaluate submissions after 10, 20, 30, and 40 epochs of relearning. Immediately after Retraining, RUV maintains a high acceptance rate consistent with our previous experiments. As relearning of the target data progresses, the evidence supporting acceptance by RUV gradually weakens, eventually leading to rejection. Therefore, RUV does not continue to accept submissions after relearning merely because the model previously underwent unlearning and exhibited representational changes. After 10 epochs, class-level target accuracy recovers from 0.013 to 0.866 while RUV acceptance decreases from 0.996 to 0.248; the final class-level/sample-level acceptance rates fall to 0.033/0.071. These results provide empirical evidence that RUV can distinguish unlearning, information-preserving updates, and relearning in the evaluated scenarios, addressing the concern that any target-specific representational change would be sufficient for acceptance. In the revised manuscript, we incorporate these experiments to explain more explicitly the basis for using representational responses as evidence for unlearning verification, as well as the empirically validated scope of their applicability.

AuthorsOpenReview ↗

Response to W1— (2)

Regarding your concern about necessity, you raise the possibility that unlearning may not produce detectable changes in at least two of RUV's evidence views. We respond in three parts.

  1. Table 8 of the revised manuscript presents the responses of different unlearning methods across multiple evidence views, showing that the response strength of each evidence view varies across methods. Together with the high acceptance rates of correct Retraining submissions in the main experiments, these results provide empirical support for RUV's ability to detect genuine unlearning in the evaluated scenarios.
  2. Different unlearning methods affect the four evidence views of the representation propagation chain to varying degrees due to differences in their underlying unlearning mechanisms. Therefore, RUV does not require all four evidence views to respond simultaneously, as illustrated by the results in Table 8 of the revised manuscript. However, a single evidence view may be affected by incidental changes. To enable mutual corroboration, we therefore require responses from at least two evidence views for unlearning verification. We further validated this design through ablation experiments, as reported in Appendix G.6 and Table 15 of the revised manuscript. Specifically, we compared the verification performance when requiring one, two, three, or four evidence views to satisfy the decision criteria. The results show that the two-view rule achieves the best overall performance in the evaluated settings. A detailed analysis is provided in our subsequent response to Q1.
  3. We acknowledge that certain special cases may exist in which RUV fails to detect unlearning. However, their relevance to the scenarios considered in this paper should be assessed in light of the conditions under which they arise. In our subsequent response to Q1, we provide a detailed analysis of the cases you raised involving unchanged or only slightly changed representations. Accordingly, we have further clarified the applicability limitations in the Discussion subsection of the revised manuscript.

Regarding Theorem D.6. We agree that D.6 cannot be directly interpreted as providing a false-acceptance guarantee against all information-preserving models. In the revised manuscript, we explicitly clarify its null hypothesis and the scope of its guarantee immediately after the theorem, avoiding the interpretation that rejecting the no-response null hypothesis is equivalent to establishing complete information removal.

Regarding Theorem D.8. In the revised manuscript, we explicitly position it as a conditional analysis of detection power: the experiments examine whether the studied unlearning methods produce these responses, while the theoretical analysis explains the conditions under which such responses can be detected when they are present.

Regarding Appendix C.4. We agree that the cost analysis of the reference-guided forgery strategy cannot rule out other information-preserving internal modifications. In the revised manuscript, we have narrowed the relevant claims and explicitly clarified the subject and assumptions of the analysis, without extending its conclusions to exclude all adaptive internal modifications.

Regarding Figure 2. We provide a detailed explanation in our subsequent response to Q1.

AuthorsOpenReview ↗

Response to W2

Thank you for your reminder. We have revisited and refined our analysis of related work. The main points are summarized as follows.

Several recent studies have also raised concerns about the reliability of output-level verification. However, most of them have not implemented concrete attacks to demonstrate practical deception [1–4], remaining largely at the conceptual level. Other studies have considered only a single unlearning scenario, either class-level or sample-level unlearning [6,7], without providing comprehensive coverage. ODR implements concrete attacks to achieve output-level deception under both class-level and sample-level unlearning requests, further revealing the manipulability of output-level verification.

Motivated by similar concerns, some studies have shifted their verification focus toward the model's internal behavior. TruVRF measures the gradient sensitivity of forgotten data [8], but requires white-box access to model parameters and backward computations, and its verification signal remains coupled with the model's output residuals. Other studies have explored representation-based verification [1–5]. However, most of these methods rely on retraining [1–3], which limits their practicality. Moreover, these methods generally address only one unlearning scenario, either class-level or sample-level unlearning [1–4]. Therefore, there remains a lack of broadly applicable verification methods that are entirely independent of retraining. Although RULER-M4 [1] does not require retraining, its authors explicitly emphasize that M4 is only a preliminary attempt and exhibits strong dataset dependence. IDI [5] already supports both class-level and sample-level unlearning and provides a retraining-free extension. However, it operates in a white-box computational setting, and its retraining-free variant still relies on surrogate reference models and requires additional training of a mutual information estimator, introducing corresponding access and computational requirements.

From a practical deployment perspective, avoiding independent retraining and reducing verification costs are important objectives. Motivated by these considerations, we propose RUV. RUV does not rely on an independently retrained model or require the construction of surrogate unlearning reference models. It only requires access to the forward representations of the original and unlearned models through a gray-box interface, without requiring white-box access to model parameters, gradients, or training logs. Meanwhile, RUV supports both class-level and sample-level unlearning requests. It makes verification decisions by combining four complementary evidence views, matched retained-data controls, and joint permutation calibration, without requiring additional training of a mutual information estimator.

A detailed analysis of the two related studies you mentioned [5,9] is provided in our subsequent response to Q3.

[1] Georgina Cosma et al., RULER: Representation-Level Verification of Machine Unlearning. arXiv preprint arXiv:2605.27569, 2026. https://doi.org/10.48550/arXiv.2605.27569.

[2] Teresa Pui Yee Yong et al., Erased, but Not Gone: Output Forgetting Is Not True Forgetting. arXiv preprint arXiv:2606.25001, 2026. https://doi.org/10.48550/arXiv.2606.25001.

[3] Yongwoo Kim et al., Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols. Engineering Applications of Artificial Intelligence, 167:113785, 2026. https://doi.org/10.1016/j.engappai.2026.113785.

[4] Yurim Jang et al., Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearning. Proceedings of the ACM Web Conference 2026, pages 8505–8508, 2026. https://doi.org/10.1145/3774904.3792896.

[5] Dongjae Jeon et al., An Information Theoretic Evaluation Metric for Strong Unlearning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(26):22173–22181, 2026. https://doi.org/10.1609/aaai.v40i26.39373.

[6] Yichen Gao et al., An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. Proceedings of the 29th International Conference on Artificial Intelligence and Statistics, PMLR 300:4114–4122, 2026. https://proceedings.mlr.press/v300/gao26e.html.

[7] Heng Xu et al., Really Unlearned? Verifying Machine Unlearning via Influential Sample Pairs. IEEE Transactions on Dependable and Secure Computing, 23(1):1671–1686, 2026. https://doi.org/10.1109/TDSC.2025.3620308.

[8] Chunyi Zhou et al., TruVRF: Toward Triple-Granularity Verification on Machine Unlearning. IEEE Transactions on Information Forensics and Security, 20:4844–4859, 2025. https://doi.org/10.1109/TIFS.2025.3565991.

[9] Binchi Zhang et al., Verification of Machine Unlearning is Fragile. Proceedings of the 41st International Conference on Machine Learning, PMLR 235:58717–58738, 2024. https://proceedings.mlr.press/v235/zhang24h.html.

Reviewer_FiH8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes post hoc verification in machine unlearning and questions the common reliance on output predictions. The authors show that Output-Level Deceptive Response (ODR), via ODR-Head and ODR-Gate, achieves high deception success rates against existing output verifiers while leaving representation parameters unchanged. This paper proposes Representation Unlearning Verification (RUV), a gray box verifier that audits intermediate forward representations. RUV evaluates a four-stage response through Activation Route Shift, Representation Shift , Neighborhood Persistence Gap, and Manifold Rank Drop. It uses matched retained controls to isolate target responses from model drift and combines evidence views with joint permutation calibration. Experiments on CIFAR10, CIFAR100, and FaceScrub show that RUV accepts genuine unlearning executions and rejects ODR submissions with high accuracy.

Critical strengths
  1. The paper demonstrates that output-based unlearning verifiers are vulnerable to simple manipulation. ODR-Head and ODR-Gate provide evidence that matching unlearning predictions does not guarantee representation change.

  2. The proposed RUV operates without requiring an independently retrained reference model. Combining matched retained controls with joint permutation calibration provides a practical, request-specific decision boundary without training separate verifiers.

  3. Experimental results demonstrate the superiority of the proposed method. Also, the authors run 1000 binary audits per setting across multiple seeds, report standard errors consistently, and provide detailed ablations on execution ratios and individual evidence views.

Critical weaknesses

While I found the paper interesting, specific concerns hesitate me to evaluate it as acceptable:

  1. the authors are encouraged to review more related papers and provide a detailed comparative discussion to clearly distinguish their contribution from existing work. For example, recent work such as [1] already demonstrated that unlearning baselines primarily modify the final classifier while preserving internal representations on ImageNet.

  2. In my understanding, RUV only measures whether target representations changed relative to the original model, but it does not verify whether the resulting representations match the distribution of a true retrained model. An update that merely corrupts or scrambles target features could trigger high displacement and neighborhood disruption, passing RUV without achieving true unlearning.

  3. The evaluation is restricted to small-scale classification on CIFAR10, CIFAR100, and FaceScrub using primarily ResNet backbones. Although ViT is tested, it is also confined to CIFAR-10. As demonstrated in prior studies such as [1], unlearning algorithms often exhibit severe representation collapse on large-scale datasets like ImageNet-1K, showing behaviors fundamentally different from small-scale scenarios. Large-scale benchmarks like ImageNet-1K and modern Transformer-based architectures should serve as the primary evaluation setting to demonstrate that RUV remains reliable in practical environments

  4. while the paper is motivated by the utility of a forward gray box auditor, this setup has a substantial reality gap with real-world deployments. In practice, models are typically served either as black-box APIs (such as GPT or Gemini) where intermediate forward hooks cannot be established without provider cooperation, or as open weights (such as Qwen) where full white-box access is already available. Furthermore, contemporary open-access models and services are predominantly generative models, whereas RUV is strictly restricted to closed-set classification due to its reliance on label-compatible reference sets for NPG and MRD (please let me know if i misunderstood something here). Although the high-level motivation of regulatory auditing is understandable, the proposed gray-box classifier setting is disconnected from modern deployment realities.

[1] "Are we truly forgetting? a critical re-examination of machine unlearning evaluation protocols." Engineering Applications of Artificial Intelligence.

Questions for authors
  1. How does RUV conceptually and empirically differ from recent representation-based benchmarks such as [1]? Can the authors clarify the practical trade-offs between using matched retained controls in RUV versus relying on an explicit retrained reference model?

  2. How does RUV distinguish between a model that properly unlearned target data and a model whose target representations were simply corrupted by random noise or adversarial perturbation? Does targeted representation degradation trigger false acceptances under the current four-stage metric?

  3. As observed in [1], many unlearning methods degrade significantly or collapse when scaled to ImageNet-1K and Transformer architectures. Can the authors provide evidence that RUV maintains its verification accuracy on ImageNet-1K and modern vision backbones without being misled by large-scale representation collapse?

  4. Given that modern deployed models are either black-box APIs or open-weight generative models, what specific real-world compliance scenario genuinely justifies a forward gray-box runtime for classifiers? How can an external auditor ensure that intermediate activations are not spoofed if the provider manages the execution environment?

Please let me know if I have misunderstood any part of the paper. I look forward to the authors' response and remain open to further discussion.

5 comments
AuthorsOpenReview ↗

Response to W1 and Q1

We thank the reviewer for the suggestion. We have expanded the discussion of Kim et al. (2026) [1] and additional related work in Sections 1–2 and Appendix A, with a dedicated empirical comparison in Appendix G.4 and Tables 12–13 of the revised manuscript. Specifically, Kim et al. (2026) show that favorable output-level evaluation results for certain unlearning methods can coexist with retained internal representations or representation collapse on large-scale datasets. They therefore propose a representation-level evaluation criterion using retraining as the gold standard. The main differences from our work are as follows:

  1. Motivation. We acknowledge that RUV and Kim et al. (2026) share a similar perspective, as both ultimately examine unlearning at the representation level. However, their work focuses on unlearning methods with inadequate performance, whereas we explicitly identify the unreliability of output-based verification and provide a concrete implementation, ODR, together with experimental analysis. This motivates us to examine the model’s internal computation and develop a comprehensive verification method along the representation propagation chain.
  2. Methodology. This is the most substantial difference. Kim et al. (2026) essentially compare representations directly with those of a retrained gold-standard model and combine this comparison with output-level evaluation to establish a unified evaluation criterion. RUV instead organizes four complementary evidence views along the representation formation process: activation routes, feature states, neighbor identities, and manifold support. It combines these views with matched retained-data controls for target attribution and joint permutation calibration for determining request-specific acceptance boundaries. The resulting verification decision accounts for both background drift and dependencies among the evidence views.
  3. Applicability. Kim et al. (2026) explicitly state that their approach is not applicable to sample-level unlearning. RUV defines and evaluates audit procedures for both class-level and sample-level requests, selects representations according to request granularity, and uses class-matched retained-data controls for sample-level requests. RUV also requires no independently retrained reference model.

Regarding the practical trade-offs raised in Q1, the approach of Kim et al. (2026) relies on an explicitly retrained model, which is difficult to obtain in practice. Unlearning is motivated by the need to avoid the high cost of retraining; relying on a retrained reference model for verification is therefore impractical. Avoiding this cost is also one of RUV’s design motivations. RUV obtains evidence from the forward representations of target samples and matched retained samples under the original and unlearned models, providing a basis for verification without incurring retraining costs. Accessing models to extract features is a reasonable and feasible requirement in practice, and Kim et al. (2026) also require such access. We appreciate the reviewer’s suggestion: our previous presentation may not have sufficiently explained this point, and we have clarified it in the revised manuscript. Because AGR requires a retrained reference, we do not include it in the main oracle-free baseline suite. Instead, we report a dedicated large-scale AGR comparison in Appendix G.4 and Tables 12–13.

[1] Yongwoo Kim, Sungmin Cha, and Donghyun Kim. Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols. Engineering Applications of Artificial Intelligence, 167:113785, 2026. https://doi.org/10.1016/j.engappai.2026.113785.

AuthorsOpenReview ↗

Response to W2 and Q2 — (1)

Thank you for your comments. We agree that the original manuscript did not sufficiently clarify the scope of RUV's guarantees. 1. Regarding the use of Retraining representations as a verification standard. We do not use consistency with Retraining representations as our verification criterion for two reasons. First, avoiding the high computational cost of independent Retraining is a primary motivation behind RUV and an important practical requirement. Second, although Retraining serves as an important unlearning reference, consistency with a particular Retraining model's representational distribution is not the only criterion for determining whether unlearning has occurred. Different unlearning methods employ distinct update mechanisms and optimization objectives, producing diverse representational responses. RUV establishes verification evidence from these responses without requiring an independently retrained reference model. 2. Regarding target attribution and multi-view verification. RUV does not merely measure changes in target representations relative to the original model. It also compares target responses against matched retained samples. Specifically, we compute both groups' responses across different evidence views, attribute changes to the target through their relative differences, and use joint permutation calibration to determine request-specific acceptance boundaries. **Therefore, large representational changes alone are insufficient to pass verification. ** The model must exhibit target-specific responses exceeding background changes in matched retained data in at least two evidence views. This design increases the difficulty of deception through noise perturbations or targeted representational disruption. Perturbations affecting both groups are constrained by retained-data controls, while selective target disruption must jointly satisfy multiple calibrated evidence views while preserving retained-data utility. Because internal feature propagation involves interconnected computations, such manipulation faces more constraints than output replacement and cannot be achieved simply through large feature shifts. **3. Regarding experimental evidence against false acceptance of non-unlearning updates or target representation disruption. ** We conducted static and dynamic controls to examine whether RUV incorrectly accepts non-unlearning updates or target representational disruption. Table 3 examines general training drift, wrong-target unlearning, and knowledge-preserving fine-tuning on the correct target set (D_u) using original labels. RUV achieves class-level/sample-level rejection rates of 0.989/0.947, 0.976/0.937, and 0.968/0.929, respectively. The last control modifies representations of the correct target set while maintaining target accuracy of 0.998/0.997. These results show that substantial model changes, wrong-target unlearning, and knowledge-preserving target updates are insufficient to pass RUV. We also conducted an unlearning-then-relearning experiment. Starting from a retained-only Retraining checkpoint, we resumed training on the complete original dataset (D=D_r\cup D_u) and audited the same (D_u) at each checkpoint. After 10 epochs, class-level target accuracy recovered from 0.013 to 0.866, while RUV acceptance fell from 0.996 to 0.248 and the margin shifted from 1.863 to −0.318. At the final checkpoint, class-level target accuracy and sample-level LiRA attack accuracy recovered to 0.935 and 0.632, approaching the Original values of 0.938 and 0.638. Meanwhile, RUV acceptance declined to 0.033 and 0.071, respectively, with stable retained-data utility. These results show that RUV does not continue accepting a model merely because unlearning was previously performed or its representations had changed. Instead, its decisions respond to the reappearance of target knowledge. Together, the static controls and dynamic relearning experiment demonstrate that RUV does not simply detect model changes or historical unlearning execution. It requires the submitted model to exhibit current internal responses consistent with the specified request, supported by target attribution and multi-view calibration. The revised Discussion further clarifies the applicability boundaries of these conclusions.

AuthorsOpenReview ↗

Response to W2 and Q2 — (2)

4. Regarding adaptive attacks specifically optimized against RUV's decision rules

We would like to emphasize that deliberately optimizing internal representations against RUV's decision rules constitutes a further-developed adaptive attack, which goes beyond the scope of a verification method itself. Based on our examination of existing machine unlearning verification studies, such methods are generally evaluated under explicitly defined threat models, without further constructing adaptive internal attacks specifically optimized against their complete decision rules. For example, Kim et al. (2026) did not investigate targeted manipulation of intermediate representations toward specific reference states to bypass their verification criteria. Our work follows this commonly adopted evaluation scope. At the same time, we go further by examining several types of model updates that could potentially cause false acceptance, including general drift, wrong-target unlearning, same-target knowledge-preserving updates, and relearning. We also analyze reference-guided forgery in Appendix C.4, showing that such an attack lacks economic justification. If an attacker first constructs a Retraining model or a genuinely unlearned reference model to obtain targets for representational manipulation, the attacker has already incurred the corresponding unlearning cost. The attacker must then perform additional optimization to align multi-view representations while preserving target knowledge and satisfying utility constraints. Under these conditions, directly submitting the already constructed reference model is no more costly than continuing to forge another model. Therefore, such an attack not only loses the low-cost advantage of output replacement, but also incurs a total cost exceeding that of directly performing the corresponding unlearning procedure.

5. Regarding the authenticity of extracted representations

We explicitly assume that the representation extraction interface faithfully returns the actual intermediate representations produced by the submitted model. Therefore, directly fabricating or replacing the extracted representations at the interface level falls outside the access setting considered in this paper. This assumption ensures the authenticity of the extraction channel, while changes within the model are audited through the target attribution and multi-view mechanisms described above.

The above revisions are reflected in Sections 3, 5.2, 6.3, and 7; Table 3; Appendices C.3–C.4 and G.3; and Tables 10–11 of the revised manuscript

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.