acceptodds
Under review as a conference paper at ICLR 2027

HOW ERASED IS ERASED? QUANTIFYING AND IMPROVING THE RECOVERABILITY OF CONCEPT REMOVAL IN CLIP

Abstract

Concept-erasure (“unlearning”) methods for vision-language models such as CLIP typically report success by showing that zero-shot accuracy on the erased class collapses to near zero. We argue this metric conflates two different things: information being removed from the representation versus information being merely hidden from one particular classifier (the zero-shot text-alignment rule). We introduce a cheap evaluation protocol – training a k-shot (k=5, 20, 50) linear probe directly on the post-erasure embeddings and measuring how much of the erased class it recovers – and apply it to three erasure mechanisms on frozen CLIP ViT-B/32 features: a naive single-direction baseline, an iterative multi-direction method (INLP), and a data-free, text-embedding-derived nullspace-projection method whose mechanism matches a recently published approach (Mishra et al., 2025). Across three image domains (CIFAR-100, EuroSAT, and Food101), we find that (i) every method we test leaves the erased concept substantially recoverable despite 0% zero-shot accuracy, and (ii) the data-free, text-derived method is consistently the most recoverable of the three (92–97% recall at k=50), more so than the naive image-based baseline, on all three datasets. On CIFAR-100, where class hierarchy labels are available, we show the leaked signal is not unstructured noise: attack false positives concentrate in the same coarse superclass as the erased class 9.8× more than chance for the text-derived method (measured over 100 randomly-sampled target classes), suggesting it erases a shared semanticdomain direction rather than a class-specific one. We also find that whether this method’s extra recoverability comes with an extra utility cost is entangled with how strong the zero-shot baseline is to begin with: the two datasets with a wellseparated zero-shot baseline (CIFAR-100, 61.7%; Food101, 84.3%) both show the text-derived method costing the most accuracy on retained classes, while the one dataset with a weak baseline (EuroSAT, 40.1%) shows the opposite – which we argue is an artifact of that weak baseline destabilizing the metric, not a genuine reversal of the phenomenon. We propose Ensemble Subspace Projection (ESP), a bootstrap-aggregated multi-direction erasure operator, which reduces attack recall relative to the classical iterative baseline at equal removed-dimension budget on every dataset we test (by 1–9 recall points), for free on CIFAR-100 but at a small utility cost on the other two; no method we test achieves what we would consider genuine erasure. All of the above uses CLIP ViT-B/32; we additionally confirm our two central findings – NSP-text is the most recoverable and the most utility-costly method – hold essentially unchanged on the 2.8× larger ViT-L/14 backbone. Finally, extending from object classes to a demographic attribute on FairFace, we find naive erasure of a binary gender attribute leaves recoverability that swings 8.9× across race groups (7.7–68.6% recall at k=50) even though every group is erased and attacked identically – a class-level mechanism producing the same kind of group-uneven outcome documented for retraining-based forgetting, and a caution that fairness properties of an erasure method cannot be read off its object-classification behavior alone; tellingly, ESP – our own defense – leaves this disparity just as wide (if not wider) even where it reduces average attack recall elsewhere. We discuss the weak-baseline pitfall and other open problems for future work on evaluating unlearning in vision-language models.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Machine Unlearning, Concept Erasure, Vision-Language Models, CLIP, Representation Recoverability, Linear Probing, Subspace Projection, Fairness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_YcgtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The author points out that after erasing a concept, CLIP zero sample classification cannot recognize the class, which does not mean that there is no relevant information in it. So, using a small number of labeled samples, the linear classifier was retrained on the erased representation to check if the class could still be recognized. The paper compares methods such as one-way projection and proposes ESP to remove multiple directions at once. The experiment verified the influence of population attributes and group differences.

Critical strengths
  1. The problem is relatively clear.
  2. Low evaluation cost and easy reproducibility.
  3. The main comparison covers a wide range of categories.
  4. Honest reporting of negative results.
Critical weaknesses
  1. The abstract is too long, almost occupying the first page. It's hard to understand what the main contribution is. The statement of 0% and group differences in the abstract is inconsistent with the more limited results in the main text. The paper emphasizes the differences in population recovery after erasure, but lacks corresponding experiments before erasure.
  2. Limited innovation and lack of differentiation comparison with the latest methods.
  3. Some charts are too large and have formatting issues.
  4. The current version is not suitable for publication in ICLR.
Questions for authors
  1. Only reporting the recovery recall rate cannot determine whether the concept can truly be effectively restored.
  2. The lack of probe controls for non erased representations makes it difficult to quantify how much information has been reduced by erasing.
  3. ESP has lower advantages compared to simple unidirectional baselines and lacks comparison with the latest methods.
  4. The fairness section needs to compare the differences before and after erasure. The abstract emphasizes a difference of 8.9 times, but the main text shows multiple sub results ranging from 1.5 to 8.9 times. Can you report the changes in the differences before and after erasure, as well as the absolute difference.
  5. A large bar chart takes up a lot of space, but lacks error bars;
1 comment
Reviewer_YcgtOpenReview ↗

Guidelines on the Use of AI Tools for Manuscript Review

This review process utilized AI to translate and organize the core content of the manuscript. Subsequently, human reviewers conducted a thorough assessment and finalized the review results. Human oversight ensured the overall quality and formatting of the manuscript, and the final conclusions were determined by human judgment rather than by AI.

Reviewer_Jbf7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines whether CLIP unlearning methods indeed erase the target concept. The authors use the linear probe to investigate this and find that most methods can be easily recovered. They also propose a defense method ESP to delete the main directions.

Critical strengths

The paper studies an important problem and the observation is crucial.

Critical weaknesses

The main concern is that almost no new or deeper analysis is provided. Linear probe is a common method for investigating residual information as in [1], and the conclusion among different unlearning methods need more ways to support.

[1] Li N, Pan A, Gopal A, et al. The wmdp benchmark: Measuring and reducing malicious use with unlearning. ICML 2024.

Questions for authors
  • ESP is a defense trained using the probe, and the evaluation also uses the probe to attack it, it would be better to change the attack model (such as MLP).
  • Only three methods are included. It would be better to evaluate more clip unlearning methods, including optimization-based methods, etc.
Reviewer_fnwQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper evaluates concept erasure mechanisms in CLIP models arguing that standard evaluation metrics of zero-shot accuracy doesn't distinguish if the information was hidden rather than removed from the model. The authors introduce a linear probing protocol to quantify the recoverability of erased concepts.They find that all erasure methods tested leave concepts highly recoverable.

Critical strengths

None

Critical weaknesses
  1. Paper is very badly written, figures are too big and there is no proper figure explaining the method. Also, the abstract is very long
  2. Only few unlearning methods were tested. Methods such as CLIPErase (Yang et al, ACL 2025), Zero-shot class unlearning in clip with synthetic samples (Kravets et al, WACV 2025) and others should be tested.
  3. Other erasure methods are ignored such as LEACE (Belrose et al., 2023)
  4. Methods are compared at very different erasure levels, e.g. DAMP 34.8% on CIFAR and 76.8% on Food101 zero-shot recall. Comparisons should be made at matched zero-shot erasure levels.
Questions for authors

See weaknesses

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

The authors in this paper state that by making a concept hidden from the original classifier doesn't imply the model has actually forgotten it, and propose a way to test that claim. First, they point out a fallacy in the current erasure methods- the erasure is tested only on a zero-shot classifier, where the model might unlearn according to the classifier-specified parts only, but the actual information about the class might still be retained. This kind of reversible unlearning is already examined in other foundational models; for VLMs, they find an analogous gap through linear probe. The linear probe is a part of k-shot recoverability evaluation protocol, and it is trained on post-erasure embeddings using few labeled examples. The analysis takes three projection-based erasure operators, naive single-direction projection, text-derived data-free nullspace projection (NSP-Text), and iterative nullspace projection (INLP). Experiments on CIFAR-100, EuroSAT and Food101 show that erase-class knowledge is highly recoverable for these baselines. To reduce the recoverability, the paper proposes Ensemble Subspace Projection (ESP), which used bootstrap sampling to aggregate regression directions and removes dominant subspace through top-k SVD. ESP reduces attack recall moderately compared to INLP. The authors also evaluate a larger CLIP backbone, and acknowledge that none of the methods actually achieve complete concept erasure.

Critical strengths
  1. The problem statement and the contributions are clearly defined, the problem definition and probe itself is significant. They distinguish between two aspects: CLIP zero-shot classifier doesn't recognize target class vs post-erasure embedding space no longer has information about the target class that is recoverable.

  2. The experiments clearly show the recoverability of the traditional methods even when the zero-shot target-class recall approaches zero.

  3. The authors further analyze correlations between subspaces that were removed and estimated attacker directions. This helps in understanding the residual information.

  4. The reports are transparent, including the negative results.

Critical weaknesses
  1. The major contribution is the linear probe investigation, which uses only erase-class recall; other metrics like AUROC, precision, false-positive rate would be valuable to report to distinguish between retained and erased classes.

  2. It would be helpful to detail compute and time overhead for ESP compared to the other methods.

  3. Semantic leakage is not analyzed in original CLIP; NSP-text produces more false positives among the classes, but is that already present in the original CLIP model?

  4. The benefit of ESP is moderately higher compared to INLP. Will increasing the number of removed directions in INLP help cover that gap?

Questions for authors

Please check weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.