Research map / Adversarial robustness and security / Attacks on text and graphs
When Priors Backfire: On the Vulnerability of Unlearnable Examples to Pretraining
Zhihao Li, Gezheng Xu, Jiale Cai, Ruiyi Fang, Di Wu, Qicheng Lao, Charles Ling, Boyu Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add random-class target labels (stage 2) after hard-negative stage (ablation) | BAIT · CIFAR-100 | test accuracy (%): 35.75 -> 23.94 | Improved |
| add most-dissimilar-class target labels (stage 3) (ablation) | BAIT · CIFAR-100 | test accuracy (%): 23.94 -> 20.77 | Improved |
| add stages 2 and 3 to curriculum on CIFAR-10 (ablation) | BAIT · CIFAR-10 | test accuracy (%): 23.68 -> 14.40 | Improved |
| add stages 2 and 3 to curriculum on SVHN (ablation) | BAIT · SVHN | test accuracy (%): 21.75 -> 14.37 | Improved |
| use ImageNet-pretrained surrogate for REM (ablation) | REM · CIFAR-10 (ResNet-18) | test accuracy (%): 74.46 -> 56.08 | Improved |
| use ImageNet-pretrained surrogate for EMN (ablation) | EMN · CIFAR-10 (ResNet-18) | test accuracy (%): 61.82 -> 59.84 | No clear effect |
| use ImageNet-pretrained surrogate for TUE (ablation) | TUE · CIFAR-10 (ResNet-18) | test accuracy (%): 82.72 -> 82.65 | No clear effect |
| apply Mixup defense (ablation) | BAIT · CIFAR-10 (ImageNet-pretrained) | test accuracy (%): 14.40 -> 21.08 | Hurt |
| apply Cutout defense (ablation) | BAIT · CIFAR-10 (ImageNet-pretrained) | test accuracy (%): 14.40 -> 13.69 | No clear effect |
| apply JPEG compression quality 10 (ablation) | BAIT · CIFAR-10 (ImageNet-pretrained) | test accuracy (%): 15.91 -> 68.08 | Hurt |
| evaluate with randomly initialized backbone instead of pretrained | BAIT · CIFAR-10 ResNet-18 | test accuracy (%): 14.40 -> 10.18 | Improved |
| use larger ViT-based pretrained backbones | UE baselines · CIFAR-10 | test accuracy (%) | Hurt |
Closest papers by meaning
- How Far Are We from True Unlearnability? (ICLR 2025 · similarity 0.89)
- One for All: A Universal Generator for Concept Unlearnability via Multi-Modal Alignment (ICML 2024 · similarity 0.88)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.87)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.87)
- On the Provable Advantage of Unsupervised Pretraining (ICLR 2024 · similarity 0.86)
- Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
