acceptodds
Under review as a conference paper at ICLR 2027

Beyond Suppression: Measuring Concept Erasure Durability in Diffusion Models

Abstract

This paper first investigates a vulnerability that quantization can reverse concept erasure in diffusion models. To this end, we first quantize diffusion models edited by different concept-erasure methods, and the results demonstrate that the target concept remains suppressed after quantization. However, the quantized erased model exhibits a novel vulnerability: restoring just 5% of the edited weights to their original values can markedly increase the generation rate of the erased concept. In this paper, we refer to the ability to keep suppression under such restoration as erasure durability and evaluate it by tracking target generation across restoration budgets. This yields a recovery curve showing how target generation changes as the restoration budget increases. We summarize this curve with the half-recovery budget, defined as the minimum restoration needed to raise the target-generation rate to the midpoint between its erased-model baseline and fully restored level. Based on this, we construct DuraBench, a durability-based benchmark that distinguishes erasure methods with similar initial suppression but different resistance to weight restoration. We further propose SPARC, a post-training method that helps erased models maintain target suppression when some edited weights are restored.

Primary area:
interpretability and explainable AI
Keywords:
Concept Erasure, Robustness, Quantization, Diffusion Models

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_fwKYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the revival of suppressed concepts in diffusion models. They first show that quantization alone does not revive concepts, but restoring different percentages (also just 5%) of the weights with the largest absolute change from those in the original model does affect reemergence. They introduce a benchmark (DuraBench) that measures erasure durability across different models. They also propose a method (SPARC) that partially restores weights and learns to maintain these responses after restoration.

Critical strengths
  • The proposed benchmark is a useful tool to investigate the suppression capability of erasure models
  • SPARC shows that there is hope that the problem of reemergence can be solved
Critical weaknesses
  • The paper needs to be better written in the story it tells: the quantization framing seems disconnected from the paper's main message. The paper could be improved, for example, by focusing less on quantization, including in the abstract. The phenomenon is not new in the literature (starting from the "The illusion of unlearning" paper, which is not mentioned)
  • SPARC is evaluated on the same prompts during evaluation of the benchmark, so explicitly on the attack it is trained on
Questions for authors
  • How does the durability ranking change with random restoration?
  • Can you include proper unlearning benchmarks such as UnlearnCanvas and HUB?
Reviewer_i4GpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the durability of concept erasure in diffusion models under quantization and partial weight restoration. The authors show that quantization alone largely preserves erasure, but restoring a small fraction of the edited weights can recover erased concepts. They introduce DuraBench, a benchmark that measures erasure durability using recovery curves and half-recovery budgets. They also propose SPARC, a post-training method that improves resistance to weight restoration.

Critical strengths
  1. The paper highlights an interesting new type of potential metric to study the robustness of concept erasure methods.

  2. The related work is well presented and nicely motivates the paper.

  3. The SPARC experiments include useful ablations

Critical weaknesses
  1. The practical importance of the proposed vulnerability is unclear. Quantization alone does not undo erasure, while restoring the original weights is a rather artificial setting. In particular, the restoration procedure selects the largest edited weights using knowledge of the original model. It is unclear why this should be considered a meaningful vulnerability, rather than an analysis tool.

  2. Section 5 is not sufficiently motivated and is difficult to follow. Much of the section reads as a collection of experimental observations without a clear explanation of why they matter. In particular, the motivation for SPARC could be explained better.

  3. The evaluation of SPARC is limited. An important question remains unanswered: does SPARC improve robustness to other forms of concept recovery, or only to the specific weight-restoration procedure? Also, much of the analysis focuses on RECE, and it is unclear how general the findings are.

Questions for authors
  1. Is there a realistic attack scenario that motivates selectively restoring the largest edited weights to their original values?

  2. Does SPARC improve robustness against other concept-recovery approaches, such as adversarial prompting or concept inversion?

Reviewer_ojz7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the durability of concept erasure in diffusion models under partial weight restoration. It first observes that the evaluated erasure methods largely retain suppression after INT8 quantization, whereas restoring a small fraction of edited weights can substantially recover the erased concept. The authors introduce DuraBench, which evaluates 14 erased checkpoints using recovery curves and a half-recovery budget. Through controlled edit decomposition, the paper further explores how additional weight changes influence restoration sensitivity. Motivated by this analysis, SPARC performs post-training response matching before and after partial restoration to improve suppression durability.

Critical strengths

The paper highlights a meaningful distinction between initial concept suppression and its persistence under subsequent weight modifications. DuraBench provides an interpretable evaluation protocol, and the comparison across 14 checkpoints demonstrates that similar initial suppression can conceal substantially different restoration sensitivity. Reporting absolute restored-weight counts alongside restoration fractions is also useful for interpreting differences across editing methods. The controlled-edit experiments offer an interesting hypothesis about how additional edits alter restoration selection, while the SPARC ablation provides evidence that restoration-response matching contributes beyond preserving initial suppression. The appendices clearly document several assumptions and limitations, including the restricted quantization model and the non-monotonicity of recovery curves.

Critical weaknesses

(1)SPARC is trained on the benchmark’s ten target prompts and evaluated using those same prompts. Since the training objective explicitly matches responses to their CLIP embeddings, the reported gains may reflect adaptation to this small prompt set rather than robust suppression of the underlying concept. Evaluation on disjoint prompts, paraphrases, compositional descriptions, and independently selected seeds is necessary. The controlled subspace construction also uses five prompts from the target list, further limiting the independence of the mechanistic evaluation. (2)Increasing the orthogonal component simultaneously changes the total edit norm, initial suppression, and the number of quantization-changing candidates. Consequently, lower recovery cannot be attributed solely to reduced restoration overlap with the proposed key component. Normalization by the initial and fully restored rates does not remove these confounders. Moreover, orthogonality holds only for a four-dimensional subspace derived from five prompts, and two additional donor models fail the initial screening criterion. Matched-norm and matched-candidate-count controls, together with held-out prompt evaluation, are needed to strengthen the mechanistic interpretation.

Questions for authors

Please refer to the weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.