acceptodds
Under review as a conference paper at ICLR 2027

BARRIER: Bounded Activation Regions for Robust Information Erasure

Abstract

Machine unlearning aims to remove targeted concepts from a trained model while preserving the rest of its knowledge. Central challenge of this setting is that effective and robust erasure requires extensive parameter updates, which can unintentionally alter representations that should be retained. As a result, existing methods often trade erasure strength for preservation, due to the lack of formal guarantees on the protection of neutral concepts. We propose BARRIER (Bounded Activation Regions for Robust Information Erasure), a method that enables more intensive unlearning by driving updates within an identified activation space control region, where target erasure can be performed with limited collateral degradation. Using interval arithmetic, we obtain a closed-form bound on the worst-case representation change over protected regions and use it as a knowledge preservation objective. We provide a formal analysis of this protection and its effect on the functional drift. BARRIER is principled, architecture-agnostic, and compatible with existing erasure objectives. Empirical evaluations demonstrate that BARRIER achieves competitive performance across classification and generative settings, including notable gains in some cases, while maintaining strong robustness against adversarial recovery attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, generative models, diffusion models, interval arithmetics, low-rank projections, information erasure

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hWdqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the challenge in machine unlearning of simultaneously erasing a target concept and preserving remaining knowledge, BARRIER proposes shifting the protection mechanism from parameter space to activation space. It performs principal component analysis on activations associated with the unlearning set at selected layers to identify the directions of representational change; within this space, it uses quantiles to define an "unlearning region" where aggressive updates are permitted, while establishing "protected regions" on either side. Using interval arithmetic, it derives a closed-form upper bound for the worst-case change in layer responses within these regions, formulating this as a differentiable regularization penalty added to the standard erasure loss, while updating only a small subset of layers. The paper demonstrates that this penalty precisely controls response changes within protected regions—maintaining validity outside these regions (albeit with bounds that loosen with distance)—and reveals the trade-off between the scope of protection and the allowable intensity of erasure. Experiments show that BARRIER significantly outperforms baselines in class erasure for diffusion models and preserves accuracy in image classification with minimal parameter updates; it exhibits robust resistance to recovery attacks without adversarial training, yet it lags behind or fails in text-to-image generation quality, class erasure for image classification, and random sample unlearning. Furthermore, its theoretical guarantees apply only to the regularization term and single-layer responses rather than the model as a whole, and ablation studies fail to clearly demonstrate that the interval-based protection term is the primary driver of retention capability.

Critical strengths
  1. The protection term is closed-form and differentiable; it introduces no additional trainable parameters and requires no changes to the network architecture, allowing for seamless integration with any erasure loss.

  2. The paper demonstrates that the protection margin simultaneously determines two factors: the tightness of the boundary surrounding the region, and the erasure intensity. These two factors operate in opposite directions, with the weighting parameter serving only to select the operating point without altering the fundamental structure of the trade-off. This constitutes the paper's most informative theoretical contribution.

  3. In diffusion model class erasure, retention accuracy reaches approximately 99%, whereas all baselines reporting this metric collapse to around 14%, while generation quality remains comparable to the retraining-based reference.

Critical weaknesses
  1. In Appendix F.1, Table 11, the variant that keeps only the projection and disables the interval terms reaches UA 99.0 and RA 99.4. Both are higher than the full method at 95.20 and 98.62. Only FID is worse. That undercuts the claim that the interval protection term is the source of retention.

  2. The interval bound constrains the penalty and the response of one layer. It does not constrain the trained model. It does not constrain end to end behaviour. Across layers the paper offers only a Lipschitz argument with no constants. The phrase formal guarantees on the protection of neutral concepts in the abstract and introduction reads stronger than what is actually delivered.

  3. Under random data forgetting UA is about 0.5 to 2.8 percent, which means the forget samples are still classified correctly. Almost nothing gets erased. On class erasure with Stable Diffusion on Imagenette the average is UA 79.60 and FID 9.14, well behind SalUn at 99.82 and 1.22.

  4. On Stable Diffusion NSFW, FID moves from 14.10 to 16.61 and CLIP from 31.34 to 30.39, while GLoCE and SAeUron reach the same erasure with equal or better utility. In the ablations the weight is reported as the best setting found, and the main tables never state how it was chosen. Optimistic bias is a real risk here.

  5. The interval bound is a standard result from interval analysis. The same interval activation consolidation already appeared in continual learning. Constraining the same response drift at a finite set of points already exists in closed form editing work. What is new here is the continuous region and the tradeoff it exposes. That is enough for a conference paper. It is not the first guarantee of its kind.

Questions for authors
  1. Was the SVD only row in Table 11 produced under exactly the same protocol and the same search budget for the weight? It beats the full method on both UA and RA, yet the text calls it the worse tradeoff. Please report the 2×2 ablation.

  2. In Tables 1, 2, 4 and 9, is the weight fixed or selected on a validation set? Did the baselines and BARRIER receive the same tuning budget?

  3. Can you provide at least one piece of end to end empirical evidence, for example the relation between per layer response drift and final retention accuracy, that connects the single layer bound to the protection of neutral concepts?

  4. Does UA near zero on random data forgetting mean the method does not apply to that task? Either state the applicability condition in the abstract, or give a construction that makes the task work.

  5. Can you report quantitative retention metrics on prompt sets unrelated to the target concept, rather than FID plus qualitative figures alone?

  6. Are the numbers in Tables 1, 2 and 3 from a single run? Please give means and variances over multiple seeds.

  7. Is the base erasure rate of 3.15 percent in Table 3 the same metric as the NudeNet value of 17 in Table 1 under a different convention?

3 comments
AuthorsOpenReview ↗

Response to the review

We thank the reviewer for their review and insightful feedback. Firstly, we refer the reviewer to the general reponse where we address the major weaknesses. We respond to the reamaining concerns below.

W1/Q1: Each Table 11 row is the best of its own sweep, and the full-method row was Table 9’s single run. Both tables now report the seeded mean at λ = 2 (UA 93.5, RA 99.0, FID 16.0). The missing cell, λ = 0, is clearly worse: UA 75.2, RA 84.5, FID 30.2 (G5, App. L.1). On RA, “SVD only” (99.4) and the full method are both within 0.7 points of the unchanged model (99.7, G3), they differ in UA (99.0 vs 93.5) and FID (17.76 vs 16.0).

W2/Q3: The bound is per layer, yet only the protected layers change and the rest of the network is frozen, so the output change is bounded by the per-layer changes times the Lipschitz constants of the frozen layers that follow (we do not compute them). The end-to-end evidence you asked for (App. L.4, 36 checkpoints): across λ, retain drift tracks FID (ρ = +0.78) and RA (ρ = −0.76), and removing the penalty (λ = 0, one seed) raises retain drift 3.7-fold and worsens all metrics. Holding λ fixed, the penalty still tracks retain drift (within-λ ρ = +0.76) and FID (+0.41, p = 0.045), and forget drift tracks UA (+0.80), while retain drift alone predicts FID and RA only weakly between seeds (+0.21 and −0.15, not significant). So drift and quality move together under λ and the penalty measures the change it limits, we do not claim drift alone explains quality. We improved the descriptions in the manuscript to be more precise. For additional details, please see G1 in the general response.

W3/Q4: We agree that BARRIER is especially beneficial for unlearning classes and concepts, which were our main targets. We improved the abstract to state this directly. Random data forgetting is reported for completeness, there its results are similar to SEMU’s, while retraining itself has UA of only 5.24% / 7.91%. The reason is measurable (App. L.2): under class erasure 0% of retain activations fall inside the forget box, under random forgetting 9–11% do, the same rate as forget activations, so no region can be changed without affecting retained data. Forgetting only part of a class does the same for the samples of that class that stay in the retain set. For additional details, please see G2 in the general response. We agree that we trail SalUn on SD Imagenette.

W4/Q2: Agreed on SD utility versus GLoCE and SAeUron, the text now says so. These results use only the erasure objective plus our penalty, with no retain loss or other mechanism. For the additional details regarding λ selection and tuning, please see G4 in the general response.

W5: We agree that interval bounds are standard and that InTAct came first; we do not suggest otherwise. BARRIER differs from InTAct in two ways that both matter (G6): it inverts the roles, leaving the forget box free and protecting the boxes beside it, and it operates in the SVD subspace of the forget activations ("IA only", the closest variant to InTAct, loses RA and FID). As you note, the continuous protected region for unlearning and the trade-off it exposes are new. The penalty is closed-form, computed once from activations, adds no trainable parameters, and can be added to any erasure objective.

Q5: We follow the I2P protocol: FID and CLIP are computed on COCO prompts unrelated to the target.

Q6: Tables 1–3 are single runs, now stated. Table 4 has 3 runs.

Q7: Same checkpoint, different metrics. Table 1 counts NudeNet detections over 4,703 I2P prompts. Table 3’s “Erased” is the share of the 95 STEREO nudity prompts flagged by NudeNet without attack (3/95), following the STEREO protocol exactly. Caption clarified.

We hope that these responses address your concerns. However, if anything still remains open, or if they raised any additional questions, please let us know, and we will be pleased to answer.

Dear Authors,

Thanks for your detailed response.

Compared to existing approaches, the performance improvement offered by your method is not significant; in fact, it is even inferior to some established techniques. Although your method is applicable to a broader range of tasks, its overall performance requires further enhancement. Furthermore, I must point out a critical issue: many of the images included in your paper—for instance, those in the first column of Figure 1—would very likely be flagged as "sensitive" (NSFW) by an NSFW classifier. I suggest that you continue to improve the method's performance and, at the same time, revise the selection of images presented in the paper.

In light of the above, I will maintain my original score.

AuthorsOpenReview ↗

Response to the rebuttal comment

Thank you for the follow-up.

Images. We agree that explicit or borderline imagery does not belong in the paper, regardless of any detector’s score. We used NudeNet to filter those which should not be there. However, as pointed by the Reviewer, there is plenty of such detectors. To make no doubts about the images, we have decided to blur the first column of Figure 1.

Performance. We do not claim to outperform every method on every benchmark, and the revised paper states where BARRIER trails (GLoCE and SAeUron on Stable Diffusion utility, SalUn on Imagenette, MACE under RAB). As the Reviewer noted in the review, the contribution is a bound over continuous regions of activation space rather than at a finite set of points, and the trade-off it exposes between the protected margins and the room left for erasure. It becomes a closed-form penalty with no additional parameters that can be added to any erasure objective. Where it helps most:

Recovery attacks: among methods without adversarial training, the lowest attack success rate under CCE (13.7% vs 46.3% for the next best, RECE) and UD (37.9% vs 41.9%). DDPM: at comparable erasure, better FID than unprotected or restricted updates (16.0–18.3 vs 25.3–40.1, App. L.1), with retain accuracy within 0.7 points of the unchanged model. Flux: the lowest NudeNet total among compared methods (171), on par with the best baseline (173), at the base model’s CLIP score.

Other concerns (the Table 11 protocol, end-to-end evidence, the choice of λ, random forgetting, the Table 3 metric) concerned soundness and the precision of our claims; the revision addresses each of them. We would be grateful to know whether any of these remain open.

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

The paper proposes a new way to protect retained knowledge during unlearning. Instead of a retain loss, it looks at the activations of the forget concept at a chosen layer, draws a box around them in a low dimensional subspace, and uses interval arithmetic to bound how much the layer output can change just outside that box. That bound becomes a penalty on the weight update. The idea is that you can then erase aggressively inside the box while the rest of the model is formally protected. The method is tested on NSFW erasure in Stable Diffusion and Flux, on adversarial recovery attacks, and on class unlearning for classifiers and DDPM.

Critical strengths

(1) The preservation penalty itself is a clean piece of engineering. Everything it needs, the subspace, the box and the margins, is computed once from activations before training, so the penalty is a closed form expression in the weight update with no trainable parts and little cost. The authors are also clear that the resulting guarantee is about the training objective, not about the trained checkpoint, which makes the paper easier to assess.

(2) The ablation appendix tests the mechanism rather than only the headline numbers. It varies the margin shape, the percentile that defines the box, and the individual terms of the penalty, which is more than most unlearning papers do.

Critical weaknesses

(1) The whole point of the paper is that a formal bound on what happens outside the forget region lets you erase harder inside it without hurting the rest of the model. The utility numbers, however, say the opposite. On Stable Diffusion the method has the worst FID of all compared methods and a low CLIP score, and GLoCE in particular beats it on erasure and on both utility metrics at the same time. On Flux the FID increase is among the largest in the table. The adversarial attack table uses the same checkpoint, so it carries the same bad FID, and there MACE, which also has no adversarial training, turns out to be far more robust on one of the attacks. The authors explain the FID as the price of an aggressive operating point, but an aggressive operating point is exactly what the bound was supposed to make safe, and in any case FID is applied the same way to every baseline. I think the root of this is visible in Section 3.2, where the paper itself says the control region should be understood as the region where the update is allowed to act rather than as a description of the concept. If the box does not actually contain the concept, then forget activations outside the box are protected instead of erased, and conversely retain activations inside the box get no guarantee at all. Seen that way, the random forgetting experiments, where almost nothing ends up unlearned, are exactly what you would expect.

(2) Beyond the utility problem, the ablations show that the interval geometry is not even what produces the results. Replacing the per direction margins with a single uniform margin changes nothing, the percentile that defines the box has no effect, a simple width floor matches the full method, and the variant that drops the interval terms entirely does slightly better. What is left after all of that is a norm penalty on the weight update, weighted by the singular values of the forget activations. That kind of penalty is close to the null space constraints used in continual learning and to the covariance preservation terms used in model editing, yet none of those are cited or compared. Relatedly, the paper says it was inspired by InTAct but never says how it differs from InTAct beyond leaving the inside of the box unconstrained, and no experiment shows that leaving it unconstrained matters.

(3) There is one table that at first looks like a decisive win for preservation. In the DDPM class unlearning table the method keeps retain class accuracy near 99 percent while the baselines sit around 14 percent, and the caption accordingly says the method is an order of magnitude better. That comparison does not hold up, though. The baseline rows are copied from the SEMU paper, and in that paper the column in question is labelled TA rather than RA, is never defined for the generative setting, and is near chance for every method even when FID is near retrain. So the baselines' 14 percent and the method's 99 percent are not the same quantity, and the order of magnitude claim compares two different things.

Questions for authors

(1) Please state concretely how BARRIER differs from InTAct and show an experiment that isolates that difference.

(2) The forget box is percentile trimmed but the outer bounds that define the margins are not. Why the asymmetry, and how sensitive are the margins to a single extreme retain activation? The appendix finding that the percentile does not matter suggests the outer range dominates everything.

(3) For the case without retain data, Section 3.2 defines a fallback scalar margin, but no value for it appears anywhere, including the hyperparameter table. Which experiments used it, and with what value?

(4) In the DDPM table, what quantity is the baseline column, how was your retain accuracy measured, and are the two comparable? If not, please re-evaluate the baselines with the same classifier or remove the claim.

(5) Several methods that beat BARRIER on utility in the Stable Diffusion table are absent from the Flux table. Why?

3 comments
AuthorsOpenReview ↗

Response to the review

We thank the reviewer for their review and insightful feedback. Firstly, we refer the reviewer to the general reponse where we address the major weaknesses. We respond to the reamaining concerns below.

W1 (utility): Agreed that the bound does not make an aggressive operating point free: it does not cover retain activations inside the forget box. Two points of context. FID registers any global shift in appearance, not only damage to benign content. And these results come from the plain erasure objective plus our penalty alone, with no retain loss, adversarial training or extra modules. SD and Flux are presented as competitive erasure with attack robustness. MACE is more robust under RAB.

W1 (box): Measured at the protected layer of ResNet-18 (G2, App. L.2). Under class erasure no retain activation lies in the box. Under random forgetting 9–11% do, the same rate as forget activations, so the box cannot tell them apart, as you predicted. Forgetting 50% or 10% of a class does the same for the samples of that class kept in the retain set (17–20% in the box, against 17–19% of the forgotten ones). As Section 3.2 says, the box marks where the update may act, not the concept itself: each of its 32 intervals spans the central 90% of the forget activations (82.3% per direction in Stable Diffusion, App. D), 18% lie inside on all 32 jointly, and class-wise forgetting at this layer still reaches 100% UA (Table 6).

W2: The penalty matters: λ = 0 gives RA 84.5% and FID 30.2, restricting the update to the forget subspace without the penalty gives RA 79.0–87.4% and FID 37.2–40.1 (App. L.1), and the penalty tracks the measured retain drift even between seeds at a fixed λ (within-λ ρ = +0.76, App. L.4). Against InTAct, both differences matter (G6). Against null-space constraints: at the protected layer the forget subspace holds 94–99.5% of retain variance (App. L.2), so freezing the response on retain directions would leave little room for erasure, whereas the box leaves the forget region free and, under class erasure, contains no retain activation. Uniform margins matching per-direction margins shows the method does not depend on finely estimated margins. Related work added.

W3/Q4: You were right that the two columns were not comparable. Re-running the evaluation, we found an error in the code behind SEMU's reported TA values, so the ~14% are not valid retain accuracies G3. We removed the baseline values. The order-of-magnitude claim is withdrawn.

Q1: Two differences, both necessary. Roles are inverted: InTAct protects the box, BARRIER leaves it free and protects the two boxes beside it. App. C.1 shows that protecting the forget box leaves little room for erasure. BARRIER also projects onto the forget SVD subspace. Interval constraints without the projection (“IA only”, the closest variant to InTAct) lose RA (92.7%) and FID (22.47). Please, see G6 in the general response for the additional details.

Q2: The two bounds play different roles. The forget bounds decide where the update is free, so we trim them to keep outliers from enlarging that region. The region to protect is the whole complement of the forget box, which is unbounded, and interval arithmetic needs finite boxes. The outer bounds are only a finite stand-in for it and need only match the scale of the activations (Table 13, the “no remain” row of Table 9). We take the observed range, or γ without retain data.

Q3: Used only for Table 9’s “no remain” row, with γ = 20. It shows that in some cases the retain set can be omitted entirely, which is desirable.

Q5: GLoCE, SAeUron, CPE and ScaPe have no official Flux results or implementation, so they appear only in the Stable Diffusion table. We would be glad to add other Flux baselines if you have specific ones in mind.

We hope that these responses address your concerns. However, if anything still remains open, or if they raised any additional questions, please let us know, and we will be pleased to answer.

Thank you for the detailed response and the new experiments, but my main concern remains. The lambda sweep shows the penalty as a whole is needed, but not that the interval construction is. The SVD only variant still has higher UA and the same RA as the full method, and the difference is 1.8 FID points between a single run and a seeded mean. Your new class embedding only baseline reaches the same RA with no protection at all, so on DDPM the penalty buys generation quality and nothing else, and with the SEMU values gone Table 9 no longer shows BARRIER preserving retained classes better than any existing method.

On utility you agree that GLoCE and SAeUron are better on Stable Diffusion and that the bound does not cover retain activations inside the box. Presenting the results as competitive erasure with attack robustness is fair, but it is a different claim from the one the paper makes, and the new frontier plots are for DDPM only. The reading of Section 3.2 is now honest, and it means random and partial class forgetting fail by construction rather than by tuning.

The paper is cleaner than before but not stronger in its main claim, so I keep my score.

AuthorsOpenReview ↗

Response to the Reviewer comment

Thank you for the careful follow-up. We ran the experiment your main concern calls for.

Interval terms vs. SVD only. You were right that Table 11 compared a single run with a seeded mean. We repeated the full λ sweep for the SVD-only variant (interval terms off, L_mean and L_res kept), 3 seeds per λ (new App. L.5). At no λ does it approach the full method. At comparable erasure its best setting (λ = 10) reaches UA 95.9 ± 1.5, RA 86.8 ± 0.6 and FID 22.4 ± 0.2, against 97.6 ± 1.0, 99.2 ± 0.2 and 18.3 for BARRIER at λ = 1; its best RA over the whole sweep is 87.8%. Without the interval terms, the projection terms protect little more than no penalty at all (λ = 0: RA 84.5%, FID 30.2). The interval terms are therefore what preserves the retained classes on DDPM, not only generation quality. The earlier single-run SVD-only row (RA 99.4%) was “lucky seed”; we replace it with the seeded results.

Comparison with existing methods. With the reported SEMU values withdrawn, Table 9 no longer supports a claim of better preservation than existing methods, and we do not make it; re-evaluating the baselines was not possible within the rebuttal period.

Claim of the paper. We agree that “competitive erasure with attack robustness” is narrower than the original framing. Beyond the abstract, we revised the contributions and the Figure 1 caption so that the paper as a whole makes the narrower claim. Random and partial-class forgetting are outside the scope by construction, as the abstract now states. The frontier plots remain DDPM-only; a Stable Diffusion sweep was not feasible in the rebuttal window.

Thank you again; your review has helped us a lot to improve our work.

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

This paper proposes BARRIER, a method for machine unlearning, where the goal is to remove the influence of a subset of the training data, defined as the forget set, from the model. Ideally, the resulting model should behave like a model that has never seen the forget set during training.

BARRIER is applied to representations before selected linear layers or affine transformations. It uses PCA on the forget-set representations to identify a lower-dimensional subspace, then uses the 5th and 95th percentiles of the projected coordinates to define a central forget region and two adjacent boxes (or hyper-rectangles) where changes to the model's response are penalized.

The authors use interval arithmetic to derive an upper bound on these changes and incorporate it as a geometric regularization term alongside existing unlearning objectives, such as random relabeling for classification and prediction matching under alternative conditioning for generative models. Moreover, they introduce additional penalties on changes along the orthogonal complement of the selected PCA subspace and on the response change at the forget-set mean.

Finally, they provide theoretical results regarding the preservation bounds and evaluate their method across image classification and generation tasks.

Critical strengths
  • The main contribution of the paper is a relatively simple regularization term that is added to existing unlearning objectives. The regularizer has a closed-form expression, does not require additional trainable parameters, and can be applied to different architectures.

  • The authors provide a large number of experiments across different architectures and tasks, including image classification, generation, adversarial robustness, and ablation studies.

Critical weaknesses

w1) The assumptions behind the proposed geometric regularization and its generality are not sufficiently investigated. My main concern is that the PCA subspace identified from the forget set may also contain a substantial amount of information needed by the retain set. If the forget and retain representations, or the information they encode, overlap significantly within this subspace, I am not convinced that an affine transformation is expressive enough to disentangle this information and perform a localized modification for unlearning without also affecting the retain set. This suggests that the method may work best when most of the important retain information lies in the orthogonal complement of the forget subspace, so that modifying the forget subspace causes little damage to the retain set. If this is the case, it raises the question of how much of the method's success comes from the proposed geometric regularizer, rather than the separation already present in the pretrained representations. This could also explain why the method appears more successful in removing entire classes or concepts, while its results on random-data forgetting in CIFAR-10 and CIFAR-100 are considerably weaker. This is only a hypothesis, but I think it is an important one that the paper does not sufficiently investigate, particularly since general data unlearning is not necessarily equivalent to concept erasure.


w2) The authors provide a large number of experiments, but it is difficult to identify where their method actually provides a meaningful improvement over existing approaches. In many cases, BARRIER performs similarly to the baselines or improves one metric while sacrificing another. Since the method involves a trade-off between forgetting and preservation, I think the authors should vary the regularization strength (λ\lambda) and plot the performance across two competing metrics, such as forgetting accuracy versus retain accuracy or FID. Comparing these curves against similarly tuned baselines would help us understand whether BARRIER achieves Pareto-optimal points at some points and under what conditions it actually offers an advantage. Furthermore, when the performance differences between BARRIER and the baselines are small, particularly for FID and CLIP, it is unclear whether these improvements are statistically significant, since the authors do not report uncertainty estimates for the main generative experiments. The qualitative evaluation also relies on a handful of selected images, which makes it difficult to judge whether the method consistently preserves benign content better than the baselines.

Questions for authors

q1) Regarding W1, could the authors compare BARRIER with a simpler approach that restricts weight updates to the PCA subspace (say wfw_f is only allowed to change), using the same unlearning objective but without the proposed geometric regularizer? It would also be useful to measure how much of the retain-set information is captured by the forget PCA subspace. Also, could the authors consider a setting where only a fraction of one class is forgotten while the remaining examples of that class are retained? Comparing this with a retrained model would help investigate the hypothesis raised in W1.

q2) Regarding W2, could the authors vary λ\lambda and show trade-off plots against baseline, to demonstrate if BARRIER achieves Pareto-optimal points? It would also be useful to report uncertainty estimates, particularly when the differences are small in FID and CLIP.

q3) In DDPM results in Table 9, the reported retain accuracy of 98.62% is considerably higher than the approximately 14–15% reported for the baselines. However, the corresponding results in SEMU are reported as Task Accuracy (TA), rather than Retain Accuracy (RA). Are these metrics calculated exactly the same?

q4) The explanation of the CIFAR-10/100 random-forgetting results in Appendix E.1–E.2 seems inconsistent with Theorem 7. In particular, a wider forget region and narrower margins make the upper bound larger (less restrictive), while the authors use this geometry to explain why less room is available for unlearning. Furthermore, the discussion following Theorem 7 seems to interpret changes in this upper bound as changes in how much the model can actually be modified. However, the theorem only establishes an upper bound, not the actual changes achieved during optimization. Could the authors clarify these points?

q5) Theorems 1 and 2 provide bounds on changes at individual affine layers. When BARRIER is applied to multiple layers, the output of each modified layer passes through nonlinear transformations before reaching the next one. These transformations may amplify or dampen the previous changes. How do the theoretical bounds account for this effect and extend to the overall model?

3 comments
AuthorsOpenReview ↗

Response to the review

We thank the reviewer for their review and insightful feedback, especially regarding their hypothesis that BARRIER relies on separation already present in the pretrained representation. We checked it carefully and already included in the revised version of the manuscript, which, we believe, was improved significantly by this finding. Firstly, we refer the reviewer to the general reponse where we address the major weaknesses. We respond to the reamaining concerns below.

W1/Q1: Your hypothesis is right, and BARRIER is designed to use this separation of internal representations, we now say so explicitly. We ran the three analyses you asked for.

  • Retain information in the forget subspace (App. L.2). At the protected layer of ResNet-18 the 32 forget directions hold 94–99.5% of retain variance (random subspace: ~6%), so directions alone cannot separate forget from retain. Location within the subspace does: under class erasure no retain activation lies inside the box, under random forgetting 9–11% do, the same rate as forget activations.
  • Restricted updates (App. L.1). On DDPM, with the same objective and layers but no penalty, updates restricted to the forget subspace reach UA 84.8–93.8%, RA 79.0–87.4% and FID 37.2–40.1 (2 seeds, unrestricted: 75.2%, 84.5%, 30.2), against 97.6%, 99.2% and 18.3 for BARRIER at λ = 1. Separation is needed, as random forgetting shows, but not enough: restricting the update does not protect the retained classes, the penalty does.
  • Partial-class forgetting (App. L.2). When 50% or 10% of the airplanes are forgotten, the kept airplanes lie inside the box at the same rate as the forgotten ones (17–20% vs 17–19%), so for them it behaves like random forgetting, as your hypothesis implies. This is a forward-pass measurement, we did not run unlearning or retraining in this setting.

W2/Q2: The 4-seed λ sweep (G4, App. L.3) shows the trade-off. Revised paper adds the UA–FID, UA–RA and RA–FID frontiers. Matched sweeps for every baseline were not feasible, so we make no Pareto claim against them. Our aim is not to dominate every baseline but to reach competitive results with a lightweight, closed-form penalty that can be added to any erasure method.

Q3: Q3: Your question led us to an error in the evaluation behind SEMU's reported values G3; they are not comparable with our TA. We removed them. Under the corrected code the unchanged model scores 99.7% and BARRIER 99.0%.

Q4: You are right on both points. App. E.1, E.2 and E.5 are rewritten around the overlap, not Theorem 7, and the introduction and App. C.1 now read Theorem 7 as a limit on the response change obtainable per unit of penalty, not as the change achieved.

Q5: The theorems bound each protected layer. Only those layers change and the rest of the network is frozen, so the change in the model’s output is bounded by the per-layer changes times the Lipschitz constants of the layers that follow them, amplification or damping by the nonlinearities is exactly what these constants capture. We do not compute them, so this is not a numerical end-to-end guarantee. Empirically (App. L.4, 36 checkpoints): across λ, the drift of the protected layers moves with FID (ρ = +0.78) and RA (ρ = −0.76), and removing the penalty raises retain drift 3.7-fold and worsens all metrics (one seed, App. L.1). At fixed λ the penalty still tracks retain drift (within-λ ρ = +0.76) and FID (+0.41), while retain drift alone predicts FID and RA only weakly between seeds, we claim co-movement under λ, not that drift alone explains quality.

Do the overlap analysis, the restricted-update baseline and the partial-class measurement answer your hypothesis about pre-existing separation, and does the corrected reading of Theorem 7 resolve your q4? If anything remains unclear, or if they raised any additional questions, please let us know, and we will be pleased to answer.

Thank you for the detailed response and additional experiments. I appreciate the authors addressing several of my concerns.

Regarding W1/Q1, I still have a few questions

  1. If 94–96% of retain variance is captured by the forget (PCs) subspace, what is the contribution of the regularization term that protects its orthogonal complement? Moreover, I think it is still possible that the forget and retain representations occupy approximately orthogonal directions within this shared 32-dimensional subspace, so this variance alone doesn't show this. Maybe showing something like subspace distance with different ranks (Grassmann distance) can verify this. (PCA on retain set and pick the first PCs and PCA on forget set and then calculate the subspace distance between these two)
  2. If only 18% of forget representations lie inside the forget box, how does this geometric construction explain the successful removal of the remaining 82%?
  3. More importantly, I am still not convinced how an affine transformation can perform localized changes inside the forget box without substantially affecting representations outside it. The fact that no retain samples lie inside the box does not establish whether such a selective (affine) update is possible.

For W2/Q2, I appreciate the additional experiments and the trade-off plots. However, the authors now emphasize that the method is lightweight, simple, and achieves competitive performance. I think these advantages need to be demonstrated more clearly. Is it computationally cheaper or easier to tune than existing methods? The closed-form expression is for the preservation regularizer, not the unlearning procedure, and does not guarantee information removal (no statistical gurantee).

Moreover, since the authors now clarify that the method mainly targets class, it acts more similar to a concept erasure method and I think comparisons with existing linear concept-erasure methods become relevant. Given the strong existing methods in this setting, it remains unclear what practical or theoretical advantage BARRIER provides other than competitive performance. I think having competitive performance in this saturated field is not a strong contribution.

Thank you for the follow-up.

Orthogonal term. Without protection, 83% of the update’s penalty falls in the orthogonal complement (App. L.4): the update spends most of its change in directions erasure does not need. L_res closes them. On its own it protects little (App. L.5). It complements the interval terms.

Directions within the subspace. We agree, variance share does not show this. Our evidence is location: under class erasure no retain activation lies in the box. We will add principal angles between forget and retain subspaces to the final version. The 82% is outside the box. They are not in the protected regions either (those require every coordinate in a margin), so the penalty constrains them weakly (Theorem 2). Removal comes from the erasure objective on all forget samples and the penalty limits how much of that change reaches retained inputs.

Affine updates. We agree. An affine layer cannot act only inside a box, and the text no longer suggests it. The penalty makes the update selective, not local: forget drift is 4.4–5.5× retain drift (App. L.4), the margin floor alone keeps RA at 99.3% while the centre term alone drops it to 70.8% (Table 12), and without the interval terms RA stays ≤ 87.8% at any λ (App. L.5). Cost and tuning. We have no trainable parameters or auxiliary models as one SVD per protected layer, computed once; 0.05–3.41% of parameters updated, then one weight λ, stable over [1, 10] across four seeds. We make no wall-clock claim.

Removal guarantee. Agreed. Now, the limitations state that the bound concerns preservation only.

Linear concept erasure. We compare with ESC/ESC-T (SVD-subspace erasure, Table 4) and UCE (closed-form linear edit, Table 2). LEACE-style methods remove linear information from a fixed representation while BARRIER bounds a layer’s change during fine-tuning, so they are complementary. We will discuss this in related work.

Contribution. Competitive numbers alone are not the claim. The claim is a plug-in preservation objective that measurably works: removing its interval terms costs ≥ 11 RA points and ≥ 4 FID at comparable erasure, and it gives the lowest attack success among methods without adversarial training (CCE 13.7% vs 46.3%).

Thank you for these questions. They helped us to sharpen the paper’s narrative of what the geometry does and does not do. Do the seeded ablation (App. L.5) and the drift measurements (App. L.4) address your concern about what drives preservation?

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response - part 1

We thank the reviewers for their insightful reviews and questions. In the following paragraphs we respond to the most crucial weaknesses mentioned by the reviewers. We address the remaining weaknesses and questions in the detailed responses to the particular reviewers. Moreover, we improved the manuscript accordingly and uploaded the revised version with the changes marked in blue.

G1. The guarantee is per layer, not end to end.

  • Concern: The abstract promised “formal guarantees on the protection of neutral concepts”, but the theorems only bound how much one protected layer’s output can change as a training objective.
  • Response: We agree that the description should be more precise. We have the formal guarantees for the closed-form per-layer bound used in the training phase, not a guarantee for the trained model. We support end-to-end preservation empirically (App. L.4), and we changed the abstract, introduction, Section 3.1 and conclusions to use the more precised terms.
  • Why a per-layer bound is enough: In principle the output change is bounded by the per-layer changes times the Lipschitz constants of the downstream layers, but these constants are loose and we do not compute them; we therefore support end-to-end preservation empirically (App. L.4).
  • New evidence (App. L.4): For 36 DDPM checkpoints (the λ sweep and λ = 0) we passed one fixed batch of noised images through the original and the unlearned model and measured how much the output of each of the 17 protected layers changed: retain drift on retain-class inputs, forget drift on airplane inputs. λ is set by us, so the sweep is a controlled i ntervention. Among settings that still erase (UA ≥ 93.5%), a larger λ lowers retain drift and improves FID from 22.2 to 16.0–18.3, and removing the penalty (λ = 0, one seed) raises retain drift 3.7-fold and worsens UA, RA and FID together. Because λ drives all of these, we also controlled for it (partial correlation given λ, and correlation within each λ across seeds). Three relations hold at fixed λ: the penalty tracks retain drift (within-λ ρ = +0.76, p < 10⁻⁴), the penalty tracks FID (+0.41, p = 0.045), and forget drift tracks UA (+0.80, p < 10⁻⁴). Seed-to-seed differences in retain drift alone predict FID and RA only weakly (+0.21 and −0.15, not significant), partly because the seed-to-seed standard deviation of RA is below 0.2 points. We therefore claim that drift and quality move together as λ changes and that the penalty measures the change it is meant to limit, not that drift alone explains quality.

G2. Random-sample and partial-class forgetting.

  • Concern: Under random forgetting UA stays near zero. Reviewer UbmN suggests BARRIER only works when the forget data is already separated from the retain data in the representation.
  • Response: We agree with the reviewers that BARRIER is especially beneficial in the settings when we unlearn classes and concepts, which were our main targets. Whereas the random-data forgetting setting is reported for the completeness. However, even these results are similar to SEMU’s (CIFAR-10, 10%: UA 0.53 vs 0.60, TA 94.94 vs 94.22), whereas even the retraining itself has UA of only 5.24% / 7.91%.
  • New evidence (App. L.2): At the protected layer of ResNet-18 we measured where retain activations lie relative to the forget data. The 32 forget directions contain 94–96% (class erasure) and 99.5% (random forgetting) of the retain variance, against about 6% for 32 random directions, so retain data uses the same directions in both cases. What differs is location: the forget box contains no retain activation under class erasure, but 9–11% under random forgetting, the same rate as forget activations, so no region can be changed without changing retained inputs. Forgetting only 50% or 10% of the airplanes behaves like random forgetting for the airplanes kept in the retain set: 17–20% of them lie in the box, against 17–19% of the forgotten ones. Overall only 1.0–1.8% of retain activations lie in the box, because these airplanes are a small part of the retain set.

We share the responses to the rest of the major weaknesses in the next official comment due to the space limit.

AuthorsOpenReview ↗

General response - part 2

G3. DDPM retain accuracy against the baselines.

  • Concern: Table 9 compared our RA (~99%) with the baselines’ ~14%, copied from SEMU, where the column is labelled TA.
  • Response: The reviewers were right to question this. When re-running the evaluation, we found an error in the evaluation code used to produce SEMU's reported TA values: no resize of input images to 224 x 224. The reported ~14% are an artifact of this error and are not a valid measure of retain accuracy.
  • New evidence (App. L): With the same code the unchanged pretrained DDPM scores RA 99.7%, so ~99% is the model’s own level. BARRIER stays within 0.7 points of it (99.0% at λ = 2), while without the penalty RA falls to 84.5%.
  • Change: We removed the baseline TA values from Table 9 and the "order of magnitude" claim. Under the corrected code the unchanged pretrained DDPM scores RA 99.7%, so ~99% is the level of the model itself; BARRIER stays within 0.7 points of it (99.0% at λ = 2), while without the penalty RA falls to 84.5%.

G4. How λ was chosen, and the λ trade-off frontier.

  • Concern: λ was reported as “the best setting found” without seeds, so results may be optimistic.
  • Response: λ was chosen per model (on a single seed) by a coarse search over orders of magnitude, then a small finer grid, judged on UA, FID and generated samples. Baseline numbers come from the original papers, so tuning budgets can differ. We state it clearly.
  • New evidence (App. L.3): A DDPM sweep over 9 values of λ from 0.01 to 100, 4 seeds each (3 at λ = 2), traces the trade-off λ controls. For λ from 0.01 to 10, erasure stays at UA 93.5–97.6% with RA 98.5–99.3%, and FID improves from 22.2–22.4 at λ ≤ 0.1 to 16.0–18.3 at λ = 1–10 (unchanged model: RA 99.7%, FID 15.14). At λ = 30 erasure starts to fail (UA 60.6%), and at λ ≥ 50 the protected layers barely move (UA 1.7–1.8%, FID 15.2–15.3, against 2.6% and 15.14 for the unchanged model). λ = 1–10 is the knee of this frontier, and the result does not hinge on one tuned value. The revised paper adds the UA–FID, UA–RA and RA–FID frontiers.

G5. Is the protection term what preserves retained classes?

  • Concern: In Table 11 the variant without interval terms (“SVD only”) scored higher UA/RA than the full method, at a worse FID.
  • New evidence (App. L.1): With no penalty at all (λ = 0, same 17 layers and loss), UA is 75.2%, RA 84.5% and FID 30.2, worse on all three metrics than any λ from 0.01 to 10. Restricting the update to the forget subspace instead of penalizing it (the simpler baseline the reviewer UbmN proposed, λ = 0, 2 seeds) does not protect either: RA 79.0–87.4% and FID 37.2–40.1, against 99.2% and 18.3 for BARRIER at λ = 1, consistent with retain data lying in the same subspace (G2). Removing only the projection (“IA only”) also costs RA (92.7%) and FID (22.47). The penalty as a whole is needed. Updating only the class embedding without protection is a strong simple baseline: slightly higher UA and the same RA as BARRIER (98.2 / 99.2), but FID 25.3 against 16.0–18.3, so BARRIER’s advantage in the DDPM setting is generation quality at comparable erasure. “SVD only” trades FID for UA/RA.
  • Table 11: Each row was the best of its own sweep, and the full-method row was a single run. It now reports the seeded mean at λ = 2 (UA 93.5, RA 99.0, FID 16.0, 3 seeds).

G6. Related work and the difference from InTAct.

  • Response: We added InTAct, null-space continual learning, null-space unlearning and covariance-preserving editing. BARRIER differs from InTAct in two ways, and both matter. It protects the two boxes beside the forget region instead of the region itself, since App. C.1 shows that protecting the region itself leaves little room for erasure. And it works in the SVD subspace of the forget activations: without the projection (“IA only”, the closest variant to InTAct) RA drops to 92.7% and FID rises to 22.47.
AuthorsOpenReview ↗

The post rebuttal status

Dear Area Chair,

We thank the reviewers for constructive reviews. In brief:

Status. hWdq raised the score to weak accept (“your response addresses all the concerns I raised”). We addressed KSHA’s main remaining concern with a new seeded experiment (point 2), and answered UbmN’s follow-up questions on how the geometry works (point 3).

Main changes:

Claims narrowed. The theory gives a per-layer bound used in training, not a guarantee on the trained model. The abstract, introduction and Figure 1 now say so, and state the scope (classes and concepts, not random-sample forgetting). The interval terms are what preserves retained classes (KSHA). In seeded sweeps on DDPM, removing them drops retain accuracy from 99.2% to at most 87.8% and raises FID from 18.3 to at least 22.4 at comparable erasure. The earlier single-run ablation suggesting otherwise was not reproduced and has been replaced. What the geometry does (UbmN). The penalty makes updates selective rather than local: drift on forget inputs is 4.4–5.5× that on retain inputs, and the margin-weighted term alone keeps retain accuracy at 99.3%, against 70.8% for the centre term alone. Evaluation error found and fixed. Reviewers rightly questioned the DDPM baseline values taken from SEMU and we found an evaluation error behind them, removed them, and withdrew the “order of magnitude” claim. Added: multi-seed λ sweeps with trade-off plots, end-to-end drift measurements, an analysis explaining why random forgetting is out of scope, and related work (InTAct, null-space and model-editing methods).

Limitations we state openly: on Stable Diffusion, GLoCE and SAeUron achieve better utility. We provided the trade-off plots for DDPM-only. We stated the bound concerns preservation, not removal.

Contribution: a closed-form penalty, with no trainable parameters and usable with any erasure objective, that bounds response drift over continuous regions of activation space. It gives the lowest attack success rate among methods without adversarial training (CCE: 13.7% vs 46.3%).

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.