acceptodds
Withdrawn from ICLR 2027

From Unlearnable Outcomes to Decodable Shortcut Codes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unlearnable examples (UEs) add small perturbations to training images so that models trained on them perform poorly on clean data. Low clean accuracy, however, does not reveal what these models learn. We find that shallow networks fit protected training sets more easily than clean ones while generalizing worse. Moving perturbations between images then separates two behaviors: predictions either follow the label associated with the perturbation regardless of the image, or depend on the combination of perturbation and image. The first behavior offers a direct route to unlearnability, because a model that decodes the perturbation needs no semantic features to fit the protected data. We formalize this route as a decodable perturbation shortcut (DPS) and propose a low-complexity shortcut codebook (LCSC), which learns one code per class that randomly initialized networks can identify on changing images. At with crop-and-flip augmentation, LCSC reduces clean accuracy to near chance on four datasets and five architectures; at , its epoch-60 accuracy on CIFAR-10 is . Its factorized variant gives the strongest protection in eight of nine JPEG settings and at adversarial-training radii of and . Both variants are constructed tens to hundreds of times faster than TAP, SEP, and REM.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
unlearnable examples, data protection, data poisoning, shortcut learning, perturbation coding, adversarial robustness

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_g2MWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates what models learn from unlearnable examples (UEs), beyond their low accuracy on clean test data. Through shallow-network probes, perturbation-transfer experiments, and output-label remapping, the authors distinguish two observed behaviors: perturbation coding and image–perturbation coupling. Motivated by these findings, they formulate a decodable perturbation shortcut (DPS) objective and propose LCSC, which jointly optimizes class-shared perturbations and a periodically reinitialized reader, together with a factorized variant, LCSC-LR. Experiments demonstrate protection across multiple datasets and architectures, with LCSC-LR providing improved resistance to JPEG compression and small-radius adversarial training.

Critical strengths
  1. The diagnostics distinguish perturbation-driven predictions from image-dependent behavior. Label remapping shows that low clean accuracy can conceal recoverable class information.

  2. LCSC-LR outperforms the tested baselines in most JPEG settings and under small-radius adversarial training. Construction is also faster than TAP, SEP, and REM.

Critical weaknesses
  1. Output-level diagnostics do not establish whether semantic information remains in intermediate features. Table 7 also shows that removing cross-class pairing or reader resets does not weaken protection at 8/255, leaving their contributions unclear.

  2. Final accuracy can overstate protection: early stopping recovers 65.3% accuracy for LCSC at 1/255. Most countermeasure comparisons use only CIFAR-10/ResNet-18 and a single seed.

Questions for authors
  1. Can linear probes on frozen features determine whether semantic information remains at early and final checkpoints?

  2. Under matched construction budgets, when do cross-class pairing and reader resets improve over class-aligned optimization and class-wise EM?

  3. Do the JPEG and adversarial-training gains persist across multiple seeds and another dataset or architecture?

  4. Why does low-rank parameterization improve JPEG resistance? Can decoding accuracy before and after compression explain the difference between LCSC and LCSC-LR?

  5. Does LCSC remain effective when only part of the training set is protected and the learner also has access to unprotected samples?

Reviewer_2S5cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first show that shallow networks can fit protected training sets more easily than clean data while generalizing worse. Then the authors introduce a perturbation-transfer diagnostic to distinguish perturbation coding, where predictions follow the perturbation's assigned label across different images, from image–perturbation coupling. Building on this analysis, they formulate the DPS objective and propose a low-complexity shortcut codebook (LCSC), which learns one class-specific perturbation through alternating optimization of a codebook and a reader network with periodic reader reinitialization. A factorized variant, LCSC-LR, further reduces the number of codebook parameters. Experiments on CIFAR and other datasets across multiple CNN architectures show near-chance clean accuracy under an 8/2558/255 perturbation budget. The authors also evaluate very small perturbation budgets, cross-class code following, JPEG compression, adversarial training, and construction cost. They report that LCSC and LCSC-LR provide strong protection in several settings.

Critical strengths

1. Behavior-oriented analysis of unlearnable examples. The paper examines not only whether protective perturbations degrade clean accuracy, but also how models respond when perturbations are transferred across images. The proposed diagnostics help distinguish perturbation-driven predictions from image–perturbation interactions, providing a useful framework for analyzing the behavior induced by unlearnable examples.

2. Compact representation and empirical coverage. The class-wise codebook substantially reduces perturbation storage compared with per-image perturbations, while experiments across multiple datasets and CNN architectures demonstrate that the approach can achieve strong protection under the evaluated settings. The reported construction times are also considerably lower than those of several optimization-intensive baselines, including TAP, SEP, and REM.

Critical weaknesses

1. The novelty over existing coding-based unlearnable examples is insufficiently established. The central idea of using class-specific perturbations as learnable shortcuts has substantial overlap with prior work, including One-Pixel Shortcut (ICLR 2023) and Perturbation-Induced Linearization (ICLR 2026). In particular, the paper's own results show that PIL, SP, and AR already achieve nearly 100% code-following rates and within-class accuracy, while PIL achieves comparable clean-test degradation to LCSC. Although the proposed reader-codebook co-optimization provides a more explicit construction objective, the paper does not sufficiently isolate its contribution or establish a clear conceptual advantage over these existing coding-based approaches. The main contribution therefore appears incremental without stronger evidence of what the proposed optimization mechanism uniquely enables.

2. The evaluation does not sufficiently establish robust data protection under realistic countermeasures. The countermeasure evaluations are primarily conducted on CIFAR with CNN (e.g., ResNet-18). The reported protection is sensitive to the training protocol: at a perturbation budget of 1/2551/255, validation-based early stopping recovers 65.3% clean accuracy for LCSC and 55.1% for LCSC-LR. Strong JPEG compression and adversarial training with radii of at least 8/2558/255 also substantially weaken the protection. Furthermore, the SP and AR baselines use native perturbations that are not matched to the nominal L∞L_\infty budget, and their low-budget variants are obtained by rescaling existing perturbations rather than regenerating them under matched constraints. These limitations weaken the reliability of the comparative claims and leave the effectiveness of LCSC against adaptive training and preprocessing strategies insufficiently established.

Questions for authors
  1. The reader and codebook are jointly optimized to minimize the decoding loss, which raises the concern that the learned codes may overfit to the specific reader architecture or its current parameterization. Periodically reinitializing the reader changes its parameters but retains the same architecture, and thus may not fully address architecture-specific co-adaptation. Although Appendix K evaluates a freshly initialized ResNet-18 on frozen codes, it remains unclear whether the codes are equally decodable by independently trained readers with different architectures. What is the authors’ rationale for expecting periodic reinitialization to prevent such co-adaptation, and what evidence supports the architecture-independent decodability of the learned codes?

  2. Class-specific perturbations and shortcut-based unlearnable examples have been explored in prior work, including PIL, SP, and AR. Indeed, the paper reports that these methods already achieve nearly 100% code-following rates and within-class accuracy. Given this overlap, what is the key conceptual insight that distinguishes LCSC from existing approaches? Beyond making perturbation decodability an explicit optimization objective and alternating between reader and codebook updates, which aspect of the proposed method represents a substantive advance over prior coding-based methods? A clearer explanation of the underlying insight and evidence isolating the contribution of the proposed optimization strategy would help establish the paper’s novelty.

  3. The construction uses only a ResNet-18 reader, and the victim evaluation primarily covers CNN architectures. Could the authors evaluate diverse readers, including CNNs and ViTs, and test the resulting codes on ViT-based classifiers and vision-language models such as CLIP? This would clarify whether the method generalizes beyond the architectures currently considered.

Reviewer_fmfxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies unlearnable examples, which add small perturbations to training images so that models trained on them perform poorly on clean data. The authors move perturbations between images and test whether predictions follow the perturbation's label or the image. They then propose LCSC, which learns one perturbation code per class. A small network, called the reader, learns to identify codes on changing images and is periodically reset, encouraging codes that different models can learn. Experiments cover four datasets and five architectures. At a budget of 8/255, trained models achieve near-chance clean accuracy; a factorized variant improves protection.

Critical strengths
  1. I appreciate the experiments that move perturbations between images and relabel model outputs. They reveal what the learner uses to predict, which clean accuracy alone misses. The image-content and independent-reader controls strengthen this analysis.

  2. The 8/255 evaluation is broad: dense LCSC reduces accuracy to approximately chance across 20 dataset–architecture combinations. The factorized version better withstands moderate JPEG compression and small-radius adversarial training.

  3. Sharing one code across a class reduces storage and construction cost. The paper clearly reports where early stopping and stronger countermeasures recover accuracy, helping readers understand the protection's limits.

Critical weaknesses
  1. The ablations do not explain why the design choices are needed. In Table 7, class-aligned pairing and removing reader resets give 10.1% and 9.9% accuracy, versus 10.6% for the full method, all with 100% code-following.

  2. Figure 2 is too complex to learn useful information about the proposed pipeline. The conceptualization of “code” is confusing: in my understanding, the key is to learn a class-specific perturbation. “Reader, Codebook, Release, Factorized Codework” are also confusing. As a paper is written for humans to read, these non-readable abstractions, in my assessment, are clear evidence of rejection.

Questions for authors

See weaknesses

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

Related papers

Discussion (0)

Sign in to comment.