acceptodds
Under review as a conference paper at ICLR 2027

ATR: Shadow-Free, Retrain-Free Evaluation of Machine Unlearning

Abstract

Machine-unlearning metrics such as TOFU Forget Quality and privacy-leakage scores do not measure forgetting in absolute terms. Each compares a signal on the forget set, such as its loss or truth ratio, to a reference (a retrained model, the retain set, or a held-out population) and reads the result against a fixed threshold. We show this reading is unreliable. On TOFU, running one method at five random seeds, with method, epoch, and data held fixed, yields Forget Quality values spanning three orders of magnitude, so a threshold verdict, and any ranking built on it, largely reflects the seed. We trace this instability to the metric's -value read-out. We then propose ATR (Artifact-Trace Reduction), an evaluator that trains no shadow models and no retrained oracle. ATR reuses the unlearning run one already performs: it reduces each example's unlearning trajectory to a compact descriptor and scores a method by a two-sample distance between the forget set's descriptors and an explicitly chosen reference. On CIFAR-10 image classifiers and a 1.5B language model on TOFU, ATR exactly reproduces the method ranking of a per-example shadow-model membership attack, the trusted but expensive gold standard (Kendall ), at roughly 7% of its cost. At 7B, ATR's ranking never flips across seeds, whereas Forget Quality's flips on a third of seed-pairs, and ATR's cross-seed spread is three to nine times tighter. Two limits apply. The score's orientation depends on the regime and must be fixed against a gold standard or calibration in new settings. And the membership-attack validation is at 1.5B, so transfer to 7B is argued rather than measured.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Unlearning, Privacy Preserving, Membership Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rpzJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the machine unlearning evaluation method and propose ATR (Artifact-Trace Reduction). Authors highlight the seed sensitivity of another evaluation methods. With ATR, which summarizes per-example unlearning trajectories into descriptors and measures the MMD between the forget set and reference descriptors. The descriptors combine loss, logit margin, gradient norm, and activation norm through six summaries. They report the results on the image classification and language model settings, demonstrating improved cross-seed stability and agreement with membership-based rankings in the evaluated settings.

Critical strengths
  • The motivation is important and clearly presented. Cross-seed evaluation stability must be emphasized.
  • ATR is intuitive and relatively simple.
  • In particular, the advantage of ATR in terms of computational cost appears to be very significant.
  • Experiments are well organized, these type of controlled analyses help explain the roles of the different components rather than only focusing on an overall improvement.
Critical weaknesses

**Major **

  • The presentation of the paper should be improved. The placement of several figures and tables interrupts the reading flow.
  • Authors argue that the direction of this indicator varies depending on the experimental setup. Doesn't this raise a fundamental concern about its generalizability as an evaluation metric? Although authors provide explanations for this behavior, but it is questionable why they do not address problems that could be solved with just minor improvements in metric design. Could authors further clarify this limitation?
  • Is it really "retrain-free"? Although the ATR score itself can be computed without a retrained model, determining its appropriate direction in a novel unlearning setting appears to require additional calibration against a gold-standard (retrain) reference. Without such calibration, it is unclear how ATR can be reliably applied to new settings. This somewhat limits its generalizability.

Minor

  • For the statements on 7B models, authors should more clearly distinguish experimentally validated results from claims based on indirect evidence. Especially, a valid membership-based ranking is not directly established at 7B scale.
  • Setup section can be moved into the Experiments section. Wouldn't it be better to cover them together in one section rather than as separate ones?
  • Calling the evaluation protocol "ATR-OOD" is somewhat misleading, given that the benchmark primarily relies on held-out in-distribution samples rather than genuinely OOD data. How about "ATR-HO" instead?
Questions for authors

See the weaknesses.

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

This paper investigates the reliability of existing machine unlearning evaluation metrics and proposes ATR (Artifact-Trace Reduction), a shadow-free and retrain-free evaluator for ranking unlearning methods. ATR records four per-example signals during unlearning, summarizes their optimization trajectories into 24-dimensional descriptors, and computes a kernel two-sample distance between the forget set and a chosen reference set.

Critical strengths
  1. The paper identifies a meaningful flaw in the standard interpretation of Forget Quality: its KS-test pp-value does not directly measure forgetting magnitude and can yield highly unstable comparisons. The experiments convincingly demonstrate how this instability distorts practical method rankings, making this analysis directly relevant to unlearning benchmarking.
  2. By reusing signals already collected during the unlearning process, ATR offers an inexpensive alternative to shadow-model membership inference attacks. Its alignment with membership-based rankings at smaller scales is encouraging, and the ablation studies effectively disentangle the respective contributions of smooth distance estimation and trajectory information. I also appreciate the authors' candid discussion of limitations, particularly regarding the unsuccessful validation of the 7B membership gold standard.
Critical weaknesses
  1. The paper's core validity claim relies on exact agreement with membership-based rankings across a relatively small set of methods and configurations. At the 1.5B scale, much of this agreement is driven by a few clearly separated membership levels, whereas the advantage of the full trajectory descriptor over simpler alternatives hinges largely on correctly ordering the retrain-proxy and NPO pair.

    More importantly, the headline 7B experiments establish cross-seed ranking stability, but not ranking correctness. Appendix D.10 reveals that the relative ATR ordering between the retrain proxy and NPO shifts across epochs, including a direct reversal from the 1.5B ordering under the matched two-epoch configuration. While this does not prove that ATR fails at 7B, it complicates any direct extrapolation from the 1.5B validation. Furthermore, the attempted 7B membership attack is dismissed as a gold standard because the gradient-difference retrain proxy falls outside its calibration regime. This leaves a critical ambiguity as to whether the failure lies with the membership evaluator itself or with the retrain proxy's inability to reproduce true forgetting behavior. Consequently, the reported stability of ATR at 7B cannot yet be interpreted as evidence of faithful unlearning evaluation at that scale.

  2. ATR measures a distributional distance between trajectory-derived descriptors, whereas the fundamental goal of unlearning concerns the behavior of the final model relative to retraining from scratch without the forget set. In principle, two unlearning procedures could converge to the exact same final model via distinct optimization paths. Their actual residual membership leakage would be identical, yet their ATR scores could diverge due to differences in intermediate losses, margins, gradients, and activations. This raises a fundamental question about what ATR actually measures: final-model forgetting, or artifacts of the optimization dynamics.

    Additionally, the directional orientation of ATR reverses between image classification and language modeling tasks. Its raw distance is therefore not intrinsically interpretable as "more" or "less" forgetting without external calibration, and the proposed oracle-free orientation mechanisms remain insufficiently validated. While these issues do not invalidate ATR as a potentially useful empirical heuristic, they substantially limit its interpretation as a general-purpose measure of unlearning.

Questions for authors
  1. Can the authors provide stronger, independent evidence of ranking faithfulness at the 7B scale?
  2. How sensitive is ATR to the optimization trajectory when the final model parameters (or outputs) are held fixed?
  3. Can the directional orientation of ATR be reliably determined without a gold standard when deployed in a previously unseen regime?
  4. Can the authors clarify the uncertainty analysis and resolve the apparent discrepancy in the cadence results?
Reviewer_6x43OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to propose a stable, reliable and efficient metric for machine unlearning, as indicated in the paper, the previous methods are not stable or expensive. The proposed method built up the metric by probing the trace within the unlearning process. The experiments are tested under image classification and generative language model.

Critical strengths
  1. How to quantify an unlearning method is important topic as it can drive the research direction in many different ways, and the paper demonstrates that some metrics could show high variance when simply re-run the methods with different seeds.
Critical weaknesses
  1. The presentation of paper needs to be improved, the clarity of the paper is unclear, and the flow of the paper is not smooth. E.g.,
  • In problem and notation section, why call a held-out set OOD, OOD has another layer of meaning, the held-out set can simply be the test set, which is never seen but still in distribution. And then the authors say "the OOD suffix names the reference slot, this default reference is held-out in distribution data, not literal OOD", if so, why call it OOD? And then the OOD term is used through whole paper.
  • The first block of Algorithm 1 is inconsistent with line 148, where the subindex of a is the unlearning steps not index of collector scalars.
  • At line 220, when enumerating the baseline methods, it does not mention is it for image classification or language model, which causing the confusion.
  • GradDiff is called the "retrain proxy".
  1. As metric to evaluate the unlearning method, it requires the unlearning trace, which could be impractical as sometimes, only the last checkpoint is stored, the paper does not study the effects on how many available checkpoints.
Questions for authors
  1. The 1.5B gold is validated by requiring GradDiff to score near chance, and at 7B it scores 0.70. Do the 1.5B conclusions hold under a validity check that does not presuppose any method's outcome, for example one anchored on the true retrain oracle you trained? Separately, how stable is the gold itself across target-model seeds (U-LiRA on images, your attack at 1.5B)?

  2. The authors claim that the proposed ATR is cheap; what if the users constraint the cost of U-LiRA to the similar regime, will the ATR still be better?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment

To all reviewers and the AC: what changed since our first responses (8 Nov).

Reviewer rpzJ's follow-up made us re-check our cost numbers against the job logs and test the low-cost setting the paper recommends. This gave new results. It also showed that some numbers in the paper and in our first responses are wrong. This comment lists both, once. No ranking averaged over seeds changes.

What the paper claims now

  1. Forget Quality ranks methods differently from seed to seed, because of its p-value readout. On the same 7B runs, ATR's ranking does not change.
  2. ATR needs no shadow model and no retrained model.
  3. ATR ranks methods as the membership gold standard does. At 1.5B this holds for all 12 method pairs the gold standard can separate. At 7B it holds for all 5, averaged over seeds, on four methods.
  4. ATR's orientation (whether a higher score means more forgotten) can be set without a gold standard, within stated limits.

Before and after

TopicSubmitted paperNow
Ranking at 7Bargued, not measuredmeasured on four methods: 5 of 5 pairs averaged over seeds, 14 of 16 single runs
Orientationset per regime against a gold standardset by a rule that needs none; right in 4 of 4 regimes, one of them an independent test; gold standard as fallback
What ATR needsthe full unlearning traceat 1.5B the original and final models are enough: 12 of 12 pairs averaged over seeds, 15 of 16 single runs
Costabout 7% and 0.6% of a shadow attackratios withdrawn; at 1.5B, 0.03 GPU-h for six methods against 1.05 for the smallest shadow attack that works
Seed stability at 7Branking never flipstrue for the stability runs (Table 1); in the gold-standard runs NPO is not seed-stable

Withdrawn or corrected

  1. Cost (§5.5, abstract; our answers to 6x43 Q2 and W-e; AC note).
    • Withdrawn: the ratios (7%, 0.6%, and our re-priced 0.9%); the 3.8-8.0% overhead; §5.5's ~5 GPU-h at 1.5B, which we cannot reconstruct; and the claim that a shadow attack's cost grows with the number of methods.
    • The paper's low-cost setting (two recordings, 100 examples per set) fails at 1.5B. It gets all 12 pairs right in only 34 of 50 random subsamples.
    • With all examples, recording takes 17.6 ± 2.9 s per run. That is a 12-28% overhead, above Appendix D.8's ≤10%.
    • The small shadow attack costs 1.55 GPU-h, not 4.3. With 4 shadow models per side (2.60 GPU-h) it gets all 12 pairs right every time.
    • Our 16-of-16 single-run result at 1.5B recorded every training step. That cost about 2.0 GPU-h for six methods.
  2. Averages and single runs (xAJT Q1, Q2; 6x43 Q1, W-e). The agreement numbers in our first responses are averages over seeds. The single-run counts are in the table. We wrote "fixed in writing before any run". The rule was fixed before the two added 7B methods were run. By then we had seen the 1.5B results and the first three 7B methods.
  3. Orientation (xAJT Q3). We said no rule works without a gold standard. That is true for the three rules described there, not for the fourth.
  4. Scale of the membership score. 0.5 is the midpoint between a model trained on the forget set and a true retrain. It does not mean "retrained". GradDiff at 1.5B is 48% of the way to a retrain. Affected: §2, Appendix A.2, Tables C.5, C.6 and D.9, §5.1, Appendices C.4, D.3 and D.14, limitation (ii) and the abstract. "Retrain proxy" becomes GradDiff.
  5. Trace length (xAJT Q4; 6x43 W-e). Limitation (iv)'s +0.83 is for the final model alone, not for the original and final models together. Table D.12's cadence row was an error Reviewer xAJT found; corrected, it is +0.97 at every cadence. Table D.20's final-versus-start statistic compares the forget set with itself, while ATR compares it with a reference. That reference is what makes the difference. With two states, the 24-number descriptor holds 8 independent numbers.
  6. Settings, and the AC note. Table C.2's batch size is wrong for NPO in the 7B gold-standard runs (batch 2, because NPO holds a reference model) and for the 7B stability runs (batch 2 with accumulation 4). In the AC note, "we have withdrawn none" is no longer true. We accept the new title: "ATR: Ranking Machine Unlearning Methods from the Unlearning Trace".

Where the detail is. The orientation rule and the cost breakdown are in our reply to Reviewer rpzJ, with the 7B account in its last comment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.