mlsanityData to 13 Oct 2026

Association Restoration Test: Revealing Restorable Shortcuts after Unlearning

arXiv 2607.05726 ↗First posted 7 Jul 2026cs.CV cs.LGcited by 0

On the map: LLM safety and security → Machine unlearning

Abstract (arXiv)

Association unlearning aims to disable learned label-attribute shortcuts while preserving task performance. Existing evaluations mainly measure output-level robustness or probe whether shortcut attributes remain readable in frozen features, but neither test determines whether a retained association remains functionally usable by the original classifier. We propose the Association Restoration Test (ART), a post-hoc diagnostic for functional shortcut restorability. ART estimates class-conditional association directions, amplifies residual components, and evaluates the modified features with the original classifier head. Across Waterbirds, CelebA, SpuCoDogs, and an ISIC timestamp-artifact extension, we show that output metrics, representation probes, and ART characterize distinct aspects of shortcut mitigation. These findings motivate restoration-aware evaluation for unlearning and shortcut-mitigation methods that target learned associations rather than individual classes or concepts.

Key facts. No citing papers in the map yet (data through 13 Oct 2026). Placed in Machine unlearning, one of 12 directions in LLM safety and security. Builds on 23 of its 29 references that are in the map. First posted on arXiv 7 Jul 2026.

Builds on

23 of 29 references.

Closest ancestors · 8

Same direction, newest first

  1. 2026An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations9 Apr
  2. 2026Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearning18 Feb
  3. 2025CLEAR: Unlearning Spurious Style-Content Associations with Contrastive LEarning with Anti-contrastive Regularization24 Jul
  4. 2025Unlearning's Blind Spots: Over-Unlearning and Prototypical Relearning Attack2 Jun
  5. 2024UnLearning from Experience to Avoid Spurious Correlations4 Sep
  6. 2023SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation19 Oct
  7. 2023Fast Machine Unlearning Without Retraining Through Selective Synaptic Dampening15 Aug
  8. 2023Towards Unbounded Machine Unlearning20 Feb

Other references · 15

  1. 2014Deep Learning Face Attributes in the Wildcited by 3,686
  2. 2020WILDS: A Benchmark of in-the-Wild Distribution Shiftscited by 1,371
  3. 2016Understanding intermediate layers using linear classifier probescited by 1,313
  4. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,242
  5. 2019Machine Unlearningcited by 1,056
  6. 2019Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networkscited by 617
  7. 2017Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC)cited by 549
  8. 2021Just Train Twice: Improving Group Robustness without Training Group Informationcited by 523
  9. 2022Last Layer Re-Training is Sufficient for Robustness to Spurious Correlationscited by 380
  10. 2020Environment Inference for Invariant Learningcited by 306
  11. 2019BCN20000: Dermoscopic Lesions in the Wildcited by 225
  12. 2022Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlationscited by 171
3 more
  1. 2020Learning from Failure: Training Debiased Classifier from Biased Classifiercited by 133
  2. 2022Nearest Class-Center Simplification through Intermediate Layerscited by 20
  3. 2023Challenges and Opportunities in Improving Worst-Group Generalization in Presence of Spurious Featurescited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning
  2. 2026Source-Free Class Relearning: Diagnosing Forgetting in Class Unlearning
  3. 2024CLIPErase: Efficient Unlearning of Visual-Textual Associations in CLIP
  4. 2026ICON decomposition: Multivariate concept auditing reveals shortcuts that neural networks rely on
  5. 2025Redirection for Erasing Memory (REM): Towards a universal unlearning method for corrupted data
  6. 2026Forgetting Similar Samples: Can Machine Unlearning Do it Better?
  7. 2025From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization
  8. 2025CAF'E: Causal Black-Box Testing of Machine Unlearning