Research map / Adversarial robustness and security / Attacks on text and graphs
Not All Poisons are Created Equal: Robust Training against 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 |
|---|---|---|---|
| Remove the 450 high-confidence poisons (ablation) | Poisoned CIFAR-10 dataset #1, supervised training from scratch · CIFAR-10 poisoned dataset #1 | poisoning success: 450/500 | No clear effect |
| Remove the 50 low-confidence poisons (ablation) | Poisoned CIFAR-10 dataset #1, supervised training from scratch · CIFAR-10 poisoned dataset #1 | poisoning success: 50/500 | Helped |
| Use self-supervised learning before the supervised phase | Poisoned CIFAR-10 training · CIFAR-10 poisoned datasets | time until poisoning | Helped |
| Use semi-supervised CRAIG with pseudo-label confidence threshold 0.2 and proportional class budgets (ablation) | CRAIG on poisoned CIFAR-10 · Poisoned CIFAR-10 datasets #1 and #6 | defense effectiveness | No clear effect |
| Drop examples with high confidence in a class other than their ground-truth class | CRAIG on poisoned CIFAR-10 dataset #6 · CIFAR-10 poisoned dataset #6 | poisoning success | Helped |
| Apply low-confidence-center dropping with a 0.6 threshold (ablation) | CRAIG on poisoned CIFAR-10 dataset #1 · CIFAR-10 poisoned dataset #1 | poisoning success | No clear effect |
| Use equal weights for selected examples after dropping low-confidence centers (ablation) | CRAIG on poisoned CIFAR-10 dataset #6 · CIFAR-10 poisoned dataset #6 | test accuracy: 91.33% -> 94.18% | Helped |
| Use equal weights for selected examples after dropping low-confidence centers (ablation) | CRAIG on poisoned CIFAR-10 dataset #6 · CIFAR-10 poisoned dataset #6 | poisoning success | Hurt |
| Lower the other-class confidence threshold from 0.8 to 0.5 (ablation) | CRAIG with low-confidence-center dropping · CIFAR-10 poisoned datasets #1 and #6 | test accuracy: 90.03% -> 75.11%; 88.51% -> 85.63%; 91.33% -> 74.96% | Hurt |
| Use smaller training subsets (ablation) | Uniform-step CRAIG · Poisoned CIFAR-10 datasets | test accuracy: around 92%; around 91%; around 85% | Mixed |
| Update the selected subset more frequently (ablation) | Uniform-step CRAIG with a 10% subset · Poisoned CIFAR-10 datasets | test accuracy: around 85%; around 86%; around 87% | Helped |
| Update the selected subset more frequently (ablation) | Uniform-step CRAIG with a 10% subset · Poisoned CIFAR-10 datasets | poisoning success: Successful poisoned datasets: 29, 35; 24, 35; 13, 29, 49 | Mixed |
Closest papers by meaning
- Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching (ICLR 2021 · similarity 0.93)
- Friendly Noise against Adversarial Noise: A Powerful Defense against Data Poisoning Attack (NeurIPS 2022 · similarity 0.91)
- On Robustness of Linear Classifiers to Targeted Data Poisoning (AAAI 2026 · similarity 0.91)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.89)
- Sharpness-Aware Data Poisoning Attack (ICLR 2024 · similarity 0.89)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.88)
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.