Research map / Adversarial robustness and security / Attacks on text and graphs
MetaPoison: Practical General-purpose Clean-label Data Poisoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use MetaPoison-crafted poisons instead of Feature Collision poisons | CIFAR-10 fine-tuning · dog-bird and frog-airplane | attack success rate | Helped |
| Remove the watermark trick from poison crafting (ablation) | MetaPoison on CIFAR-10 fine-tuning · dog-bird and frog-airplane | attack success rate | No clear effect |
| Apply Feature Collision poisons rather than MetaPoison poisons | CIFAR-10 train-from-scratch · dog-bird | attack success rate: 0% | Hurt |
| Add MetaPoison-crafted poisons at a 1% budget | CIFAR-10 models trained from scratch · dog-bird and frog-airplane across ConvNetBN, VGG13, and ResNet20 | attack success rate: 40-90% | Helped |
| Add MetaPoison-crafted poisons at a 1% budget | ResNet20 trained from scratch · CIFAR-10 dog-bird | attack success rate: 72% | Helped |
| Increase poison budget to 10% | CIFAR-10 models trained from scratch · bird-bird self-concealment | attack success rate: 100% | Helped |
| Use self-concealment poisons | CIFAR-10 models trained from scratch · airplane-airplane self-concealment | attack success rate: 95% | Helped |
| Distribute poisons across multiple classes and optimize for a chosen adversarial label | CIFAR-10 models trained from scratch · multiclass poisoning | attack success rate: 40-50% | Helped |
| Transfer poisons across architectures | CIFAR-10 models trained from scratch · architecture transferability | attack success rate | Mixed |
| Train victims with changed learning rates, batch sizes, or regularizers | MetaPoison-crafted poisons on ConvNetBN · CIFAR-10 robustness to training settings | attack success rate | Mixed |
| Train Google Cloud AutoML victims on MetaPoison-poisoned data | Google Cloud AutoML Vision · CIFAR-10 target birds at 0.5% poison budget | attack success rate: >15% | Helped |
Closest papers by meaning
- Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching (ICLR 2021 · similarity 0.88)
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.88)
- From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models (NeurIPS 2024 · similarity 0.86)
- DataElixir: Purifying Poisoned Dataset to Mitigate Backdoor Attacks via Diffusion Models (AAAI 2024 · similarity 0.86)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.86)
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences (NeurIPS 2026 · 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.