acceptodds
Withdrawn from ICLR 2027

INVARIANCE DOES NOT GUARANTEE REPAIR IN FROZEN-FEATURE DETECTORS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Removing a nuisance subspace from frozen features makes a detector exactly invariant to the removed directions. Whether that repairs the group error that motivated the removal is a different question, which we study in authenticity detectors on frozen speech and image encoders. Deleting a rank-4 emotion subspace estimated by the same procedure on acted corpora raises false rejection of surprised genuine speech from 17.3% to 33.9% on a w2v-BERT 2.0 head and lowers it on a WavLM head. What reverses the outcome on that w2v-BERT head is where the eraser is estimated. With four or five labels and matched rank, every tested eraser estimated on acted corpora raises the error, the same constructions estimated on genuine target-domain speech lower it at a cost in neutral rejection, and that speech helps more still when it trains the detector instead. The removed rank, the residual decodability of the nuisance and its mean response are three diagnostics an erasure audit can report, and none of them decides between the two outcomes. For a fixed linear head, the removed part of that response gives the direction in which the response moves but not reliably that of the error. Controlled image experiments show the limit of the mean response as well. With the audited geometry fixed, a second training association changes the group error several-fold while the mean nuisance response remains similar. Repair therefore has to be judged on target-domain decisions, for the genuine groups and generated sources at stake, not on the erasure itself.

Primary area:
interpretability and explainable AI
Keywords:
concept erasure, invariant representations, group robustness, spurious correlations, threshold calibration, deepfake detection

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

The paper studies whether removing nuisance directions from frozen representations improves group errors in authenticity detection. Speech experiments show that erasure can help or harm depending on the estimation data and the detector. Controlled image experiments show that similar mean nuisance responses can accompany substantially different group errors. The paper explains these results through changes in sample margins and recalibrated thresholds, arguing that repair must be evaluated through target-domain decisions.

Critical strengths
  1. The speech comparisons show why the data used to estimate an eraser matter for downstream errors. Comparing erasure with direct use of the same genuine speech for detector training gives the findings practical value.

  2. They provide convincing evidence that the mean nuisance response alone cannot determine group error, making the contribution broader than an isolated speech failure.

Critical weaknesses
  1. The conceptual advance is limited, since prior work has already shown that removing spurious features can harm group performance. The added value lies mainly in the controlled experiments on authenticity detectors.

  2. The primary subspace is estimated by PCA, while the null uses a label-dependent QR construction. Relabeling already formed displacement rows leaves their PCA subspace unchanged, the observed and permuted statistics need to be explicitly distinguished.

Questions for authors
  1. Relabeling the existing displacement rows leaves their PCA subspace unchanged, please clarify how the observed and permuted statistics were constructed.

  2. Using the same development threshold rule can still produce different operating costs, which affects how the reduction in genuine rejection should be interpreted.

Reviewer_B2ZwOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether removing "emotion-related" or "image-processing" directions from frozen features improves deepfake (speech and image) detection. The authors evaluate the detectors after removal, both with and without retraining the classifier. The results on speech vary with the encoder, estimation data, and speakers, while controlled image experiments show that similar average score responses can hold different error rates. The authors find that feature-removal statistics alone cannot guarantee whether detection improves for a particular group.

Critical strengths
  1. The experiments cover several removal methods, detectors, and speaker splits, making the results more informative than a single failure example.
  2. Comparing different estimation datasets helps explain when feature removal helps and when it hurts.
Critical weaknesses
  1. The paper is unnecessarily hard to read. Terms such as “eraser,” “guardedness,” and “mean response” obscure the question of whether removal reduces detection errors. The paper needs substantial rewriting before submission to other venues. For example, the starting sentence of the Conclusion – “Invariance to a nuisance subspace is a property of the representation a head sees; the error a group suffers is a property of scores, a calibrated threshold and the samples near it; and the two are joined by the detector, the eraser and the evaluation population. “ is filled with unnecessary jargon that does not add any new information and, in fact, makes it hard to understand. This appears throughout the paper, which is why I suggest the authors to seriously consider rewriting the whole paper before resubmission.
  2. The main finding is already well known: removing unwanted features can hurt accuracy and affect groups differently. The paper does not clearly explain what new understanding the detector experiments add (Khani and Liang, 2021).
Questions for authors
  1. What specific finding goes beyond prior work on the harm caused by removing unwanted features?
  2. The headline projection increases error in 36 of 168 English-speaker splits. How should the readers interpret the unusually large reference example relative to the full results?
Reviewer_VF46OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper reports extensive results in experiments on the topic identifying synthetic speech / images, and whether erasing nuisance feature subspaces can improve the accuracy. The experiments on speech and images are similar, and the paper focused on speech, which I shall summarize below.

Their baseline recipe is the following:

  1. Take a pre-trained encoder (wav2vec, or wavlm)
  2. Use the following collection of training data to train a classification head on features extracted with the pre-trained encoder
    • ASVspoof 2019: neutral real / syntheic
    • EmoSpoof-TTS: emotional synthetic only
  3. Use the following collection of data to estimate a "eraser" projection with the hope of removing nuisance subspaces
    • RAVDESS
    • CREMA-D
  4. Optional: retrain the classifier head after erasing the learned subspaces
  5. Evaluate on real / synthetic surprised and neutral speech ("surprised" is an emotional that is not included in the training data)

Outcome:

  • (Figure 1(a)) Heads from both encoders do a terrible job on surprised speech compared to neutral speech
  • (Lines 155-161) Erasing the nuisance subspaces makes wav2vec go from bad to worse; wavlm from very very bad to still very bad.
  • (Tables 2 & 3) But if we use in-domain data (estimate the eraser from training speakers from EmoSpoof-TTS instead; or train the classifier also with real EmoSpoof-TTS data), the models will do very well on surprised speech.
  • The authors looked at 3 simple diagnostics metrics (line 98) but found them to be uninformative.
Critical strengths

This paper reported a large number of experimental results, which provided a comprehensive view of the angles from which the authors tried to tackle the problem at hand.

Critical weaknesses
  • Lack of novelty: This paper employ only common, basic, and simple techniques. The central idea of removing nuisance subspaces from linear classification is over 20 years old at this point (Solomonoff et al., 2005). Applying this on features extracted from powerful pre-trained audio encoders is not fundamentally different from prior art in methodology.
  • Lack of meaningful results: It is not clear to me what this paper is trying to tell us other than just a bunch of experimental results. Tables 2 and 3 show that "repair" is largely successful, once we use in-domain data for estimating the eraser, or training the classifier head. The title of this paper reports a failure of "repair". But it appears that failure is more of a result of poor generalization from out-of-domain data.
Questions for authors

None.

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

Related papers

Discussion (0)

Sign in to comment.